# patching

Published articles for patching.

This is one page of public article previews, not the complete archive. Follow Next page to continue. Summaries are not the original full articles.

## Industry Reactions to Major Outages and AI Patching Debate: Feedback Friday

DevFeed: [Industry Reactions to Major Outages and AI Patching Debate: Feedback Friday](<https://devfeed.tech/articles/industry-reactions-to-major-outages-and-ai-patching-debate-feedback-friday-146266.md>)

Original publisher: [Read original article](<https://www.securityweek.com/industry-reactions-to-major-outages-and-ai-patching-questions-feedback-friday/>)

Author: SecurityWeek News

Published: 2026-11-20T12:38:55Z

Content type: opinion

Language: en

Sources: [SecurityWeek](<https://devfeed.tech/sources/securityweek.md>)

Topics: [Resilience](<https://devfeed.tech/topics/resilience.md>), [Production Engineering](<https://devfeed.tech/topics/production-engineering.md>), [Artificial Intelligence](<https://devfeed.tech/topics/ai.md>), [ai security](<https://devfeed.tech/topics/ai-security.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [artificial-intelligence](<https://devfeed.tech/tags/artificial-intelligence.md>), [disruption](<https://devfeed.tech/tags/disruption.md>), [featured](<https://devfeed.tech/tags/featured.md>), [feedback](<https://devfeed.tech/tags/feedback.md>), [feedback-friday](<https://devfeed.tech/tags/feedback-friday.md>), [incident-response](<https://devfeed.tech/tags/incident-response.md>), [industry](<https://devfeed.tech/tags/industry.md>), [major](<https://devfeed.tech/tags/major.md>), [outage](<https://devfeed.tech/tags/outage.md>), [outages](<https://devfeed.tech/tags/outages.md>), [patching](<https://devfeed.tech/tags/patching.md>)

### AI overview

The reactions describe a tension between faster vulnerability discovery and the time organizations need to test and safely deploy fixes. Several contributors warn that rushed or broad updates can destabilize fragile, interconnected systems; others stress that the outages have no established shared cause and should not be attributed to AI without evidence. Suggested practices include prioritizing exploitable flaws, testing changes in stages, monitoring service behavior, maintaining verified rollback and recovery plans, and designing independent failure and management paths. Contributors differ on emphasis: one advocates limiting security updates to necessary changes, while another argues for risk-based prioritization and detection to help bridge patching delays. The text reports that Microsoft 365, FAA, banking, and government systems were disrupted, with officials citing distinct technical or maintenance causes.

### Source excerpt

Experts commented on whether AI-accelerated patching played a role, the need for tested rollback plans, and the fragility of interconnected systems. The post Industry Reactions to Major Outages and AI Patching Debate: Feedback Friday appeared first on SecurityWeek.

## When Does Depth Matter For In-Context Learning? Adaptive Inference in Deep Transformers

DevFeed: [When Does Depth Matter For In-Context Learning? Adaptive Inference in Deep Transformers](<https://devfeed.tech/articles/when-does-depth-matter-for-in-context-learning-adaptive-inference-in-deep-transformers-144908.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2606.16694>)

Author: Ravin Raj, Gautam Reddy

Published: 2026-11-20T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>), [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [Transformers](<https://devfeed.tech/topics/transformers.md>), [Machine learning](<https://devfeed.tech/topics/machine-learning.md>), [Inference](<https://devfeed.tech/topics/inference.md>)

Tags: [adaptive](<https://devfeed.tech/tags/adaptive.md>), [context](<https://devfeed.tech/tags/context.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [distributed-inference](<https://devfeed.tech/tags/distributed-inference.md>), [gaussian](<https://devfeed.tech/tags/gaussian.md>), [in-context-learning](<https://devfeed.tech/tags/in-context-learning.md>), [inference](<https://devfeed.tech/tags/inference.md>), [learning](<https://devfeed.tech/tags/learning.md>), [linear-attention](<https://devfeed.tech/tags/linear-attention.md>), [linear-regression](<https://devfeed.tech/tags/linear-regression.md>), [patching](<https://devfeed.tech/tags/patching.md>), [physics-app-ph](<https://devfeed.tech/tags/physics-app-ph.md>), [q-bio-nc](<https://devfeed.tech/tags/q-bio-nc.md>), [routing](<https://devfeed.tech/tags/routing.md>), [structure](<https://devfeed.tech/tags/structure.md>), [transformers](<https://devfeed.tech/tags/transformers.md>)

### AI overview

For in-context linear regression, the theory predicts that a single transformer layer without feedforward blocks is optimal for Gaussian priors, while deeper transformers outperform it for non-Gaussian, tree-like priors. The authors propose that successive layers use intermediate "function vectors" to infer a latent context variable at finer scales, enabling adaptive information routing. Trained linear attention transformers reproduce the theory's quantitative predictions, and causal key-patching experiments support a role for intermediate function vectors in that routing. The results suggest depth and feedforward blocks are advantageous when latent variables have hierarchical structure.

### Source excerpt

arXiv:2606.16694v3 Announce Type: replace Abstract: Transformers perform computations through many successive attention and feedforward blocks, allowing them to learn complex correlations between a large collection of coupled variables. When does stacking successive attention-feedforward computations over many layers provide a computational advantage over a single transformer block? We address this question by examining in-context learning in generalized linear attention transformers. We first introduce a general theory of distributed inference in such transformers, subject to constraints on communication and depth. We show that such systems can exploit internal representations (`function vectors') to infer a latent context variable at increasingly finer scales over its layers. For an in-context linear regression task, the theory predicts that while one-layer transformers without feedforward blocks are optimal for Gaussian priors over the context variable, multi-layer transformers are superior for non-Gaussian, tree-like priors. Trained linear attention transformers reproduce quantitative predictions from the theory. Using causal key-patching experiments, we verify that function vectors in intermediate layers mediate adaptive routing of information. Our results suggest that depth and feedforward blocks enable transformers to implement adaptive inference, and this is advantageous when the distribution over latent variables has hierarchical structure.

## The internet needs a nap (and you do, too)

DevFeed: [The internet needs a nap (and you do, too)](<https://devfeed.tech/articles/the-internet-needs-a-nap-and-you-do-too-144581.md>)

Original publisher: [Read original article](<https://blog.talosintelligence.com/the-internet-needs-a-nap-and-you-do-too/>)

Author: William Largent

Published: 2026-11-19T19:00:00Z

Content type: opinion

Language: en

Sources: [Cisco Talos Blog](<https://devfeed.tech/sources/cisco-talos-blog.md>)

Topics: [Security](<https://devfeed.tech/topics/security.md>), [network security](<https://devfeed.tech/topics/network-security.md>), [observability](<https://devfeed.tech/topics/observability.md>), [alert triage](<https://devfeed.tech/topics/alert-triage.md>), [it modernization](<https://devfeed.tech/topics/it-modernization.md>), [Resilience](<https://devfeed.tech/topics/resilience.md>)

Tags: [000](<https://devfeed.tech/tags/000.md>), [1-5](<https://devfeed.tech/tags/1-5.md>), [17](<https://devfeed.tech/tags/17.md>), [2](<https://devfeed.tech/tags/2.md>), [2026](<https://devfeed.tech/tags/2026.md>), [22](<https://devfeed.tech/tags/22.md>), [about](<https://devfeed.tech/tags/about.md>), [accountability](<https://devfeed.tech/tags/accountability.md>), [actor](<https://devfeed.tech/tags/actor.md>), [admin](<https://devfeed.tech/tags/admin.md>), [after](<https://devfeed.tech/tags/after.md>), [agencies](<https://devfeed.tech/tags/agencies.md>), [alert](<https://devfeed.tech/tags/alert.md>), [all](<https://devfeed.tech/tags/all.md>), [internet](<https://devfeed.tech/tags/internet.md>), [news](<https://devfeed.tech/tags/news.md>), [over](<https://devfeed.tech/tags/over.md>), [patching](<https://devfeed.tech/tags/patching.md>), [stop](<https://devfeed.tech/tags/stop.md>), [takes](<https://devfeed.tech/tags/takes.md>), [threat-source-newsletter](<https://devfeed.tech/tags/threat-source-newsletter.md>)

### AI overview

The newsletter urges security teams to make time for steady maintenance amid a constant stream of incidents: inventory systems and owners, identify unsupported or hard-to-patch applications, keep systems updated, and learn normal network activity so anomalies are easier to spot. It stresses that patching alone is not a realistic answer and that people need time to assess alerts. Its featured alert is Java4Shell (CVE-2026-114938), an OpenJDK flaw described as enabling unauthenticated remote code execution on Java 17.0.6 and later on x86-64 CPUs with AVX-512 VBMI. No exploitation had been observed at publication. The text recommends applying the update; if that is not immediately possible, disabling AVX-512 at JVM startup with `-XX:UseAVX=2` is a mitigation for x86-64 only, and ARM systems may fail to start with that option. It also recommends the blog's Snort rules and Splunk detections.

### Source excerpt

Bill takes a break from the non-stop breaking news to promote prioritizing environmental awareness and team stability over reactive patching.

## Java4Shell vulnerability (CVE-2026-114938) assessment

DevFeed: [Java4Shell vulnerability (CVE-2026-114938) assessment](<https://devfeed.tech/articles/java4shell-vulnerability-cve-2026-114938-assessment-144121.md>)

Original publisher: [Read original article](<https://www.databricks.com/blog/java4shell-vulnerability-cve-2026-114938-assessment>)

Author: Fermín J. Serna

Published: 2026-11-19T17:00:00Z

Content type: news

Language: en

Sources: [Databricks](<https://devfeed.tech/sources/databricks.md>)

Topics: [cve remediation](<https://devfeed.tech/topics/cve-remediation.md>)

Tags: [assessment](<https://devfeed.tech/tags/assessment.md>), [cve](<https://devfeed.tech/tags/cve.md>), [cve-2026-114938](<https://devfeed.tech/tags/cve-2026-114938.md>), [disclosure](<https://devfeed.tech/tags/disclosure.md>), [java4shell](<https://devfeed.tech/tags/java4shell.md>), [mitigation](<https://devfeed.tech/tags/mitigation.md>), [openjdk](<https://devfeed.tech/tags/openjdk.md>), [openjdk-vulnerability](<https://devfeed.tech/tags/openjdk-vulnerability.md>), [patched](<https://devfeed.tech/tags/patched.md>), [patching](<https://devfeed.tech/tags/patching.md>), [performance](<https://devfeed.tech/tags/performance.md>), [platform](<https://devfeed.tech/tags/platform.md>), [products](<https://devfeed.tech/tags/products.md>), [products-services](<https://devfeed.tech/tags/products-services.md>), [public](<https://devfeed.tech/tags/public.md>), [runtime](<https://devfeed.tech/tags/runtime.md>), [runtimes](<https://devfeed.tech/tags/runtimes.md>), [serverless](<https://devfeed.tech/tags/serverless.md>), [services](<https://devfeed.tech/tags/services.md>), [vulnerability](<https://devfeed.tech/tags/vulnerability.md>)

### AI overview

As of November 19 at 10 AM Pacific, Databricks said it had mitigated CVE-2026-114938 across all supported Databricks Runtime versions by disabling the vulnerable code, with no expected meaningful performance impact. To apply the mitigation quickly, it proactively restarted affected clusters. Databricks planned patched Runtime versions as a further precaution: serverless customers would receive them automatically, while classic cluster customers would need to restart; Databricks was also contacting customers using custom images. The company reported no evidence of unauthorized access to its products, services, or customers' data.

### Source excerpt

As of November 19 at 10 AM Pacific Time:Prior to Java4Shell's public disclosure (CVE-2026-114938),...

## Chain-of-Thought Hidden States Reveal Task-Solving Directions for Direct Answers

DevFeed: [Chain-of-Thought Hidden States Reveal Task-Solving Directions for Direct Answers](<https://devfeed.tech/articles/chain-of-thought-hidden-states-reveal-task-solving-directions-for-direct-answers-142812.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.24598>)

Author: Houman Mehrafarin, Ioannis Konstas

Published: 2026-11-19T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [Chain-of-thought](<https://devfeed.tech/topics/chain-of-thought.md>), [llm-reasoning](<https://devfeed.tech/topics/llm-reasoning.md>)

Tags: [chain-of-thought](<https://devfeed.tech/tags/chain-of-thought.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [direction](<https://devfeed.tech/tags/direction.md>), [hidden-states](<https://devfeed.tech/tags/hidden-states.md>), [intermediate](<https://devfeed.tech/tags/intermediate.md>), [large-language-models](<https://devfeed.tech/tags/large-language-models.md>), [linear](<https://devfeed.tech/tags/linear.md>), [llama-3-1](<https://devfeed.tech/tags/llama-3-1.md>), [patching](<https://devfeed.tech/tags/patching.md>), [performance](<https://devfeed.tech/tags/performance.md>), [qwen](<https://devfeed.tech/tags/qwen.md>), [tokens](<https://devfeed.tech/tags/tokens.md>), [train](<https://devfeed.tech/tags/train.md>)

### AI overview

The study reports that a single question-specific rotation of a direct prompt's final hidden state can improve answer accuracy beyond chain-of-thought (CoT) prompting while using about one-tenth as many tokens. The direction is built from that question's CoT hidden states, labeled "recovering" or "failed" according to whether patching them into the direct prompt produces the correct answer. Across 25 tasks, mean accuracy rose from 46% to 85% on LLaMA 3.1 8B and from 52% to 86% on Qwen 2.5 7B; CoT reached 69% and 77%, respectively. A linear probe also distinguished the two hidden-state classes. The abstract proposes using these directions in future work to train models to align their states with them.

### Source excerpt

arXiv:2611.24598v1 Announce Type: cross Abstract: Chain-of-Thought (CoT) prompting improves the performance of Large Language Models by eliciting intermediate reasoning steps. However, CoT is far from perfect: on some tasks, models perform worse than with direct-answer prompting, and on others they generate a few hundred tokens to reach the correct answer. We start by labelling each CoT hidden state as recovering if patching it into the direct prompt yields the correct answer, and failed otherwise. A linear probe shows that the model encodes these two classes differently. We then show that the representation space of the direct prompt holds a direction for each question that leads the model, in a short generation, to the correct answer more often than CoT does. Specifically, we rotate the direct prompt's last hidden state once, before generation, towards a direction built for each question from its own recovering and failed states. Across 25 tasks, these directions raise the direct prompt's mean accuracy from $46\%$ to $85\%$ on LLaMA 3.1 8B (CoT: $69\%$) and from $52\%$ to $86\%$ on Qwen 2.5 7B (CoT: $77\%$), with a tenth of CoT's tokens. CoT is thus not the shortest route to this accuracy, and these directions can be used in future work to train models to align their states with them.

## Java4Shell (CVE-2026-114938): Check Your Java Services Now

DevFeed: [Java4Shell (CVE-2026-114938): Check Your Java Services Now](<https://devfeed.tech/articles/java4shell-cve-2026-114938-check-your-java-services-now-140644.md>)

Original publisher: [Read original article](<https://blog.checkpoint.com/security/java4shell-cve-2026-114938-check-your-java-services-now/>)

Author: raphaelge@checkpoint.com

Published: 2026-11-18T09:19:33Z

Content type: news

Language: en

Sources: [Check Point Blog](<https://devfeed.tech/sources/check-point-blog.md>)

Topics: [cve remediation](<https://devfeed.tech/topics/cve-remediation.md>), [Keycloak](<https://devfeed.tech/topics/keycloak.md>), [Logging](<https://devfeed.tech/topics/logging.md>), [openjdk](<https://devfeed.tech/topics/openjdk.md>)

Tags: [2026](<https://devfeed.tech/tags/2026.md>), [amd](<https://devfeed.tech/tags/amd.md>), [avx-512-vbmi](<https://devfeed.tech/tags/avx-512-vbmi.md>), [base64](<https://devfeed.tech/tags/base64.md>), [check-point](<https://devfeed.tech/tags/check-point.md>), [code-execution](<https://devfeed.tech/tags/code-execution.md>), [cve](<https://devfeed.tech/tags/cve.md>), [cve-2026-114938](<https://devfeed.tech/tags/cve-2026-114938.md>), [identity-providers](<https://devfeed.tech/tags/identity-providers.md>), [inventory](<https://devfeed.tech/tags/inventory.md>), [java](<https://devfeed.tech/tags/java.md>), [java4shell](<https://devfeed.tech/tags/java4shell.md>), [patch](<https://devfeed.tech/tags/patch.md>), [patching](<https://devfeed.tech/tags/patching.md>), [poc](<https://devfeed.tech/tags/poc.md>), [public-poc](<https://devfeed.tech/tags/public-poc.md>), [runtimes](<https://devfeed.tech/tags/runtimes.md>), [security](<https://devfeed.tech/tags/security.md>), [services](<https://devfeed.tech/tags/services.md>), [startup](<https://devfeed.tech/tags/startup.md>), [x86-64](<https://devfeed.tech/tags/x86-64.md>)

### AI overview

Java4Shell can cause out-of-bounds writes in Java's accelerated Base64 decoder on qualifying Intel and AMD x86-64 processors; crafted requests may lead to memory disclosure or code execution, often before authentication. Exposure requires an unpatched runtime, a specific processor, and attacker-controlled Base64 reaching Java's standard decoder. Inventory each application's actual runtime, check CPU flags, and assess input reachability. Patch with a fixed runtime; prioritize identity providers and internet-facing services. If patching is delayed, restart with -XX:-UseBASE64Intrinsics, testing for performance impact; -XX:UseAVX=2 is broader and prevents Java from starting on ARM64. Check Point's IPS blocks the public PoC but is not a complete fix. The excerpt reports no confirmed exploitation as of November 18, 2026, and describes product-specific reachability and bundled-runtime caveats.

### Source excerpt

Bottom line: A single malformed Base64 string can give an attacker code execution on unpatched Java runtimes, often without logging in. Fixes were released on November 17, 2026. Inventory your Java runtimes, patch identity providers first, and use a startup flag where you cannot patch today. Check Point IPS protection for the public PoC is available now, but patching is still the only complete fix. Our full advisory has every detail. What we see The flaw: On recent Intel and AMD x86-64 processors with AVX-512 VBMI, Java's hand-tuned Base64 decoder miscalculates the output length for the padded tail of a [...] The post Java4Shell (CVE-2026-114938): Check Your Java Services Now appeared first on Check Point Blog.

## CISA Urges Immediate Patching of Java4Shell Flaw That Could Allow Remote Code Execution

DevFeed: [CISA Urges Immediate Patching of Java4Shell Flaw That Could Allow Remote Code Execution](<https://devfeed.tech/articles/cisa-urges-immediate-patching-of-java4shell-flaw-that-could-allow-remote-code-execution-140563.md>)

Original publisher: [Read original article](<https://thehackernews.com/2026/11/cisa-urges-immediate-patching-of.html>)

Author: The Hacker News

Published: 2026-11-18T08:24:23Z

Content type: news

Language: en

Sources: [Cyber Newsroom](<https://devfeed.tech/sources/cyber-newsroom.md>)

Topics: [cybersecurity and infrastructure security agency](<https://devfeed.tech/topics/cybersecurity-and-infrastructure-security-agency.md>), [openjdk](<https://devfeed.tech/topics/openjdk.md>), [Memory Safety](<https://devfeed.tech/topics/memory-safety.md>), [code security](<https://devfeed.tech/topics/code-security.md>)

Tags: [2026](<https://devfeed.tech/tags/2026.md>), [application-security](<https://devfeed.tech/tags/application-security.md>), [cisa](<https://devfeed.tech/tags/cisa.md>), [code](<https://devfeed.tech/tags/code.md>), [code-execution](<https://devfeed.tech/tags/code-execution.md>), [critical](<https://devfeed.tech/tags/critical.md>), [cve](<https://devfeed.tech/tags/cve.md>), [cve-2026-114938](<https://devfeed.tech/tags/cve-2026-114938.md>), [cybersecurity](<https://devfeed.tech/tags/cybersecurity.md>), [cybersecurity-and-infrastructure-security-agency](<https://devfeed.tech/tags/cybersecurity-and-infrastructure-security-agency.md>), [execution](<https://devfeed.tech/tags/execution.md>), [infrastructure](<https://devfeed.tech/tags/infrastructure.md>), [infrastructure-security](<https://devfeed.tech/tags/infrastructure-security.md>), [java](<https://devfeed.tech/tags/java.md>), [java4shell](<https://devfeed.tech/tags/java4shell.md>), [oracle](<https://devfeed.tech/tags/oracle.md>), [organizations](<https://devfeed.tech/tags/organizations.md>), [patch](<https://devfeed.tech/tags/patch.md>), [patching](<https://devfeed.tech/tags/patching.md>), [remote](<https://devfeed.tech/tags/remote.md>), [remote-code-execution](<https://devfeed.tech/tags/remote-code-execution.md>), [software-security](<https://devfeed.tech/tags/software-security.md>), [vulnerability](<https://devfeed.tech/tags/vulnerability.md>)

### AI overview

CISA urges prompt patching of Java4Shell (CVE-2026-114938), an out-of-bounds write in the x86-64 HotSpot JVM's Base64 decoder that can cause denial of service or, with specially crafted input, unauthenticated remote code execution under certain conditions. Oracle and OpenJDK patches are available; the excerpt reports no known exploitation in the wild. Keycloak says risk depends on an unpatched OpenJDK build, exposure of the AVX-512 VBMI CPU feature, and enabled Base64 intrinsics without the AVX2 limit `-XX:UseAVX=2`; image tags alone do not establish exploitability. CISA recommends inventorying Java runtimes, including bundled versions, and applying vendor patches. If patching is delayed, it suggests validating and limiting Base64 input and restricting unauthenticated access. The excerpt also describes Keycloak temporary mitigations pending updated images.

### Source excerpt

CISA Urges Immediate Patching of Java4Shell Flaw That Could Allow Remote Code Execution The U.S. Cybersecurity and Infrastructure Security Agency (CISA) has urged organizations to immediately patch a newly disclosed critical security flaw impacting Oracle Java. The flaw, tracked as CVE-2 ... Read more Published Date: Nov 18, 2026 (4 hours, 42 minutes ago) Vulnerabilities has been mentioned in this article. CVE-2026-114938 CVE-2026-118383

## Innovating safely with AI: Security and engineering leaders on life after the patch era

DevFeed: [Innovating safely with AI: Security and engineering leaders on life after the patch era](<https://devfeed.tech/articles/innovating-safely-with-ai-security-and-engineering-leaders-on-life-after-the-patch-era-143180.md>)

Original publisher: [Read original article](<https://www.chainguard.dev/unchained/innovating-safely-with-ai-security-and-engineering-leaders-on-life-after-the-patch-era>)

Published: 2026-11-18T00:00:00Z

Content type: news

Language: en

Sources: [Chainguard: Unchained](<https://devfeed.tech/sources/chainguard-unchained.md>)

Topics: [Security](<https://devfeed.tech/topics/security.md>), [code security](<https://devfeed.tech/topics/code-security.md>), [ai security](<https://devfeed.tech/topics/ai-security.md>), [LLM security](<https://devfeed.tech/topics/llm-security.md>), [AI software supply chain security](<https://devfeed.tech/topics/ai-software-supply-chain-security.md>), [development-process](<https://devfeed.tech/topics/development-process.md>)

Tags: [after](<https://devfeed.tech/tags/after.md>), [agents](<https://devfeed.tech/tags/agents.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-risk](<https://devfeed.tech/tags/ai-risk.md>), [ai-security](<https://devfeed.tech/tags/ai-security.md>), [attack-surface](<https://devfeed.tech/tags/attack-surface.md>), [attacks](<https://devfeed.tech/tags/attacks.md>), [automate](<https://devfeed.tech/tags/automate.md>), [blast-radius](<https://devfeed.tech/tags/blast-radius.md>), [engineering](<https://devfeed.tech/tags/engineering.md>), [patch](<https://devfeed.tech/tags/patch.md>), [patching](<https://devfeed.tech/tags/patching.md>), [secure-development](<https://devfeed.tech/tags/secure-development.md>), [security](<https://devfeed.tech/tags/security.md>), [threats](<https://devfeed.tech/tags/threats.md>)

### AI overview

AI increases the pressure on already strained vulnerability management, but the panel argues that preventing problems from entering systems can improve both security and development speed. It recommends secure defaults, deterministic controls for agents, and shared security guidance embedded in development workflows; metrics should drive remediation rather than reward favorable numbers. Practices vary by organization: Nationwide requires a golden path and takes a tiered, risk-based approach, keeping agents off the core bank for now; Trustpilot launched agents after months of preparation and excludes some areas; Depop integrates AI capabilities with security and uses agents to reduce routine toil. The panel also favors automating routine patching, focusing on outcomes over compliance paperwork, and keeping people accountable for agent-generated changes. The article reports a discussion and examples, not a universal implementation plan.

### Source excerpt

At Chainguard Assemble London, security leaders discuss AI-driven threats, the limits of traditional patching, and the future of secure development.

## How VLMs Plan What's Next

DevFeed: [How VLMs Plan What's Next](<https://devfeed.tech/articles/how-vlms-plan-what-s-next-139200.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.21600>)

Author: Rohit Gandikota, Arnab Sen Sharma, Adam Belfki, David Bau

Published: 2026-11-17T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [vlm](<https://devfeed.tech/topics/vlm.md>), [llm-reasoning](<https://devfeed.tech/topics/llm-reasoning.md>), [model architecture](<https://devfeed.tech/topics/model-architecture.md>)

Tags: [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [cs-cv](<https://devfeed.tech/tags/cs-cv.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [cursor](<https://devfeed.tech/tags/cursor.md>), [formats](<https://devfeed.tech/tags/formats.md>), [languages](<https://devfeed.tech/tags/languages.md>), [layer](<https://devfeed.tech/tags/layer.md>), [list](<https://devfeed.tech/tags/list.md>), [mechanism](<https://devfeed.tech/tags/mechanism.md>), [patching](<https://devfeed.tech/tags/patching.md>), [plan](<https://devfeed.tech/tags/plan.md>), [pointers](<https://devfeed.tech/tags/pointers.md>), [progress](<https://devfeed.tech/tags/progress.md>), [qwen3](<https://devfeed.tech/tags/qwen3.md>), [region](<https://devfeed.tech/tags/region.md>), [retrieval](<https://devfeed.tech/tags/retrieval.md>), [select](<https://devfeed.tech/tags/select.md>), [sequence](<https://devfeed.tech/tags/sequence.md>), [size](<https://devfeed.tech/tags/size.md>), [synthetic](<https://devfeed.tech/tags/synthetic.md>), [tasks](<https://devfeed.tech/tags/tasks.md>), [vlms](<https://devfeed.tech/tags/vlms.md>)

### AI overview

The paper proposes that Qwen3-VL plans the order of objects in image descriptions through four steps: binding heads connect text tokens to image regions, then use already-written answers to track progress as a cursor; a selection step picks the nearest unwritten region or stops; and gaze heads retrieve its visual concept for language heads and late MLPs to express. Causal patching on a synthetic visual retrieval task supports each step, with tests against alternative hypotheses. The authors report that the mechanism also holds across input formats, natural photographs, and languages, and that the relevant binding, relay, and gaze heads occur in models from 2B to 32B parameters. The supplied text is an abstract, so it does not provide detailed methods or results.

### Source excerpt

arXiv:2611.21600v1 Announce Type: cross Abstract: When a vision-language model describes an image, it names one object at a time and it rarely repeats itself, yet how the model plans this order is far from obvious. We characterize the mechanism a VLM uses to decide where to look before it looks. The mechanism has four main parts. First, binding heads match the relevant text tokens against the image and resolve them into pointers to the matched image region. Second, the same binding heads bind each written answer to the region it came from, and the model turns it into a progress signal: a cursor, the position in the image up to which the list has been written. Third, a select step chooses the nearest unwritten region after the cursor, or stops. Finally, gaze heads execute the chosen pointer and return a concept of what is drawn there, which language heads and the late MLPs turn into a word in the language of the question. We validate each part of this mechanism through causal patching in Qwen3-VL models on a synthetic visual retrieval task, and test our mechanisms against alternative hypotheses. We further show that the mechanism holds across input formats, natural photographs, and languages, and that its binding, relay, and gaze heads appear at every model size from 2B to 32B. This work allows us to investigate how a VLM turns a question and an image into a sequence of answers, layer by layer and token by token. More broadly, it opens a way to study how models plan over complex multi-step tasks with long rollouts. Our project is available at https://plan.baulab.info/

## Wavelet-Energy-Driven Dynamic Patching for Time Series Forecasting

DevFeed: [Wavelet-Energy-Driven Dynamic Patching for Time Series Forecasting](<https://devfeed.tech/articles/wavelet-energy-driven-dynamic-patching-for-time-series-forecasting-139155.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.21319>)

Author: Syed Hafiz Bin Salim Alsree, Stephen Roberts, Stefan Zohren

Published: 2026-11-17T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [Time Series](<https://devfeed.tech/topics/time-series.md>), [Artificial Intelligence](<https://devfeed.tech/topics/ai.md>)

Tags: [cache](<https://devfeed.tech/tags/cache.md>), [caching](<https://devfeed.tech/tags/caching.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [datasets](<https://devfeed.tech/tags/datasets.md>), [energy](<https://devfeed.tech/tags/energy.md>), [filter](<https://devfeed.tech/tags/filter.md>), [forecasting](<https://devfeed.tech/tags/forecasting.md>), [learning](<https://devfeed.tech/tags/learning.md>), [patches](<https://devfeed.tech/tags/patches.md>), [patching](<https://devfeed.tech/tags/patching.md>), [signal](<https://devfeed.tech/tags/signal.md>), [time-series](<https://devfeed.tech/tags/time-series.md>), [time-series-forecasting](<https://devfeed.tech/tags/time-series-forecasting.md>), [traffic](<https://devfeed.tech/tags/traffic.md>), [training](<https://devfeed.tech/tags/training.md>), [transformers](<https://devfeed.tech/tags/transformers.md>)

### AI overview

The reported gains from wavelet-based forecasting appear to come mainly from multi-resolution structure and the variable-length patches it creates, rather than learning the wavelet filter. The method uses a per-scale DWT energy map to merge adjacent windows with the same dominant scale. With a frozen wavelet, it beats a fixed-patch baseline on eight of nine datasets by 1.5%-7.8%, but is 1.2% worse on Traffic. Learning the filter helps on three datasets and makes Traffic 2.7% worse. Variable-length patches outperform task-tuned uniform patches on 27 of 36 settings; recomputing boundaries during training never helps. The authors recommend caching boundaries once and treating filter learning as a dataset-specific choice.

### Source excerpt

arXiv:2611.21319v1 Announce Type: new Abstract: Patch-based transformers tokenise a time series into fixed-size, uniform windows, which commits the model to a single temporal resolution even though real signals are multi-scale. The discrete wavelet transform (DWT) is a natural alternative because it decomposes a signal across scales. A growing body of work makes the wavelet filter learnable, on the assumption that a more flexible wavelet forecasts better. That assumption has rarely been tested in isolation, because a single reported improvement usually credits three things at once. These are the use of any multi-resolution decomposition, the choice of wavelet and training the filter. We introduce a forecaster that turns a learnable-DWT per-scale energy map into variable-length patches by merging adjacent windows that share a dominant scale. Three findings follow. First, most of the improvement is there before any filter is learned. A frozen wavelet beats a fixed-patch baseline on eight of the nine datasets by 1.5% to 7.8%, and on four of them it recovers all of the gain. Traffic is the one exception, where it is 1.2% worse. Second, learning the filter gives a further improvement on only some datasets. It helps on three of the nine datasets and makes Traffic 2.7% worse. Third, having variable-length patches helps, but exactly where their boundaries fall does not. A fixed-size uniform grid, even with its patch size tuned per task, is worse than our variable-length patches on 27 of 36 settings. Recomputing the boundaries from the learned filter at every step, rather than caching them once, never helps. The improvement in wavelet-based forecasting comes from the multi-resolution structure and the variable-length patches it induces, not from making the wavelet learnable. In practice, start from any valid wavelet, cache the patch boundaries once and treat training the filter as a per-dataset decision rather than a default.

## Every AKS Node You Can Patch in Place Is a Node That Can Drift

DevFeed: [Every AKS Node You Can Patch in Place Is a Node That Can Drift](<https://devfeed.tech/articles/every-aks-node-you-can-patch-in-place-is-a-node-that-can-drift-136236.md>)

Original publisher: [Read original article](<https://blog.devops.dev/aks-node-drift-azure-container-linux-immutable-os-90df033e23ac?source=rss----33f8b2d9a328---4>)

Author: Dave R | Microsoft Azure & AI MVP ☁

Published: 2026-11-16T13:25:26Z

Content type: tutorial

Language: en

Sources: [DevOps.dev](<https://devfeed.tech/sources/devops-dev.md>)

Topics: [Azure Kubernetes Service](<https://devfeed.tech/topics/azure-kubernetes-service.md>), [Linux](<https://devfeed.tech/topics/linux.md>), [Deployment Strategies](<https://devfeed.tech/topics/deployment-strategies.md>)

Tags: [2026](<https://devfeed.tech/tags/2026.md>), [aks](<https://devfeed.tech/tags/aks.md>), [azure](<https://devfeed.tech/tags/azure.md>), [cluster](<https://devfeed.tech/tags/cluster.md>), [commands](<https://devfeed.tech/tags/commands.md>), [container](<https://devfeed.tech/tags/container.md>), [cybersecurity](<https://devfeed.tech/tags/cybersecurity.md>), [devops](<https://devfeed.tech/tags/devops.md>), [guard](<https://devfeed.tech/tags/guard.md>), [image](<https://devfeed.tech/tags/image.md>), [kubernetes](<https://devfeed.tech/tags/kubernetes.md>), [layer](<https://devfeed.tech/tags/layer.md>), [linux](<https://devfeed.tech/tags/linux.md>), [node](<https://devfeed.tech/tags/node.md>), [nodepool](<https://devfeed.tech/tags/nodepool.md>), [os](<https://devfeed.tech/tags/os.md>), [patches](<https://devfeed.tech/tags/patches.md>), [patching](<https://devfeed.tech/tags/patching.md>), [pool](<https://devfeed.tech/tags/pool.md>), [secure-boot](<https://devfeed.tech/tags/secure-boot.md>), [ubuntu](<https://devfeed.tech/tags/ubuntu.md>), [update](<https://devfeed.tech/tags/update.md>), [weekly](<https://devfeed.tech/tags/weekly.md>)

### AI overview

Azure Container Linux (ACL) replaces patchable AKS node operating systems with signed, read-only images, reducing the risk of nodes drifting apart; nodes are replaced weekly instead of patched in place. The excerpt says ACL is generally available from Kubernetes 1.34 and is the documented destination for Azure Linux with OS Guard, whose patches end December 10, 2026. Existing Ubuntu or Azure Linux pools can be moved in place with Azure CLI 2.86.0 or later; AKS reimages the pool onto ACL. The excerpt does not provide the promised verification steps or operational limitations.

### Source excerpt

How Azure Container Linux locks down the node OS layer by layer, and how to move off OS Guard before its patches end December 10 Continue reading on DevOps.dev "

## OpenAI's Daybreak brings AI agents to security testing and vulnerability fixes

DevFeed: [OpenAI's Daybreak brings AI agents to security testing and vulnerability fixes](<https://devfeed.tech/articles/openai-s-daybreak-brings-ai-agents-to-security-testing-and-vulnerability-fixes-131445.md>)

Original publisher: [Read original article](<https://www.helpnetsecurity.com/2026/11/13/openai-daybreak-vulnerability-remediation-red-team-defense-factory-agents/>)

Author: Anamarija Pogorelec

Published: 2026-11-13T09:31:07Z

Content type: news

Language: en

Sources: [Help Net Security](<https://devfeed.tech/sources/help-net-security.md>)

Topics: [Application Security](<https://devfeed.tech/topics/application-security.md>), [code security](<https://devfeed.tech/topics/code-security.md>), [Agentic SOC](<https://devfeed.tech/topics/agentic-soc.md>), [bugfixes](<https://devfeed.tech/topics/bugfixes.md>), [OpenAI](<https://devfeed.tech/topics/openai.md>)

Tags: [agentic-ai](<https://devfeed.tech/tags/agentic-ai.md>), [agents](<https://devfeed.tech/tags/agents.md>), [ai](<https://devfeed.tech/tags/ai.md>), [applications](<https://devfeed.tech/tags/applications.md>), [attack](<https://devfeed.tech/tags/attack.md>), [attack-paths](<https://devfeed.tech/tags/attack-paths.md>), [code](<https://devfeed.tech/tags/code.md>), [cybersecurity](<https://devfeed.tech/tags/cybersecurity.md>), [daybreak](<https://devfeed.tech/tags/daybreak.md>), [daybreak-red](<https://devfeed.tech/tags/daybreak-red.md>), [defense](<https://devfeed.tech/tags/defense.md>), [factory](<https://devfeed.tech/tags/factory.md>), [news](<https://devfeed.tech/tags/news.md>), [openai](<https://devfeed.tech/tags/openai.md>), [patching](<https://devfeed.tech/tags/patching.md>), [penetration-testing](<https://devfeed.tech/tags/penetration-testing.md>), [red-team](<https://devfeed.tech/tags/red-team.md>), [security](<https://devfeed.tech/tags/security.md>), [security-testing](<https://devfeed.tech/tags/security-testing.md>), [software-security](<https://devfeed.tech/tags/software-security.md>), [source-code](<https://devfeed.tech/tags/source-code.md>), [testing](<https://devfeed.tech/tags/testing.md>), [vulnerability](<https://devfeed.tech/tags/vulnerability.md>), [vulnerability-management](<https://devfeed.tech/tags/vulnerability-management.md>)

### AI overview

Daybreak pairs security testing with vulnerability remediation. Its early-access Red Team uses coordinated agents to inspect application code and context, investigate attack paths, confirm vulnerabilities, and test whether patches block exploits. It can validate scanner, bug bounty, and disclosure findings; assessments use isolated environments, defined boundaries, and Guardian review to limit agent access. Defense Factory investigates findings from security tools or Red Team and proposes code fixes. Engineers review and test patches before merging, and Codex supports examining evidence and refining changes. The ChatGPT preview plugin connects to a GitHub repository: cloud scans require each user to connect a personal GitHub account and can target a default branch or selected files and folders. Teams can scan new commits, schedule recurring scans, and review pull requests before merge. The desktop app also scans local checkouts and uncommitted changes; optional Deep scans make multiple passes and use more time and tokens.

### Source excerpt

OpenAI is bringing AI agents into more of the software security process with Daybreak, pairing web application penetration testing with tools that investigate vulnerabilities and develop fixes. The two offerings, Red Team and Defense Factory, connect the work of finding security weaknesses with patching and retesting them. Red Team tests applications and verifies fixes Daybreak Red Team, offered through an early-access program, uses coordinated agents to examine application source code and context, investigate potential attack ... More -> The post OpenAI's Daybreak brings AI agents to security testing and vulnerability fixes appeared first on Help Net Security.

## Policy and Disclosure: 2026 Edition

DevFeed: [Policy and Disclosure: 2026 Edition](<https://devfeed.tech/articles/policy-and-disclosure-2026-edition-133696.md>)

Original publisher: [Read original article](<https://projectzero.google/2026/11/disclosure-policy.html>)

Author: Natalie Silvanovich

Published: 2026-11-13T08:00:00Z

Content type: release

Language: en

Sources: [Project Zero](<https://devfeed.tech/sources/project-zero.md>)

Topics: [cve remediation](<https://devfeed.tech/topics/cve-remediation.md>), [known exploitable vulnerabilities](<https://devfeed.tech/topics/known-exploitable-vulnerabilities.md>), [Threat monitoring](<https://devfeed.tech/topics/threat-monitoring.md>)

Tags: [2021](<https://devfeed.tech/tags/2021.md>), [2026](<https://devfeed.tech/tags/2026.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-assisted](<https://devfeed.tech/tags/ai-assisted.md>), [bug](<https://devfeed.tech/tags/bug.md>), [days](<https://devfeed.tech/tags/days.md>), [disclosure](<https://devfeed.tech/tags/disclosure.md>), [discovery](<https://devfeed.tech/tags/discovery.md>), [exploit](<https://devfeed.tech/tags/exploit.md>), [exploitation](<https://devfeed.tech/tags/exploitation.md>), [faster](<https://devfeed.tech/tags/faster.md>), [llms](<https://devfeed.tech/tags/llms.md>), [malicious](<https://devfeed.tech/tags/malicious.md>), [multiple](<https://devfeed.tech/tags/multiple.md>), [patch](<https://devfeed.tech/tags/patch.md>), [patches](<https://devfeed.tech/tags/patches.md>), [patching](<https://devfeed.tech/tags/patching.md>), [policy](<https://devfeed.tech/tags/policy.md>), [remediation](<https://devfeed.tech/tags/remediation.md>), [risk](<https://devfeed.tech/tags/risk.md>)

### AI overview

Project Zero is replacing its "90 + 30" disclosure policy with a fixed deadline: disclose vulnerabilities 90 days after vendor notification, then move to 60 days on February 2, 2027, and 30 days on January 3, 2028. The change responds to AI-assisted discovery and exploitation, which raises immediate risk and makes patch rollout "soak time" less meaningful. Project Zero also aims to reduce the patch gap between a fix's release and users installing it. The staged schedule reflects vendor concerns that an abrupt shift could expose unfixed bugs, distort prioritization, or discourage attempts to meet deadlines. Most special cases, including grace periods and variant rules, are removed; exceptions cover active exploitation, details already made public, or vendor request. Project Zero will still track and communicate whether bugs are fixed, but that determination will not control disclosure timing.

### Source excerpt

Since 2021, Project Zero has followed a "90 + 30" vulnerability disclosure policy, allowing vendors up to 90 days to remediate a vulnerability plus an extra 30 days "soak time" for patch rollout. Our goal was to encourage faster patching and give vendors time to deliver patches to users. The advent of AI-assisted vulnerability discovery and exploitation has changed the threat landscape such that the vulnerability remediation timeline encouraged by our disclosure policy is clearly insufficient. LLMs have increased the speed at which attackers can find and exploit vulnerabilities, alongside the risk that multiple parties, including malicious ones, discover the same bug. A vulnerability reported today presents a much higher immediate risk to users than one reported last year. In this new environment, it is crucial that vendors move towards remediating vulnerabilities faster.

## Introducing Project Theseus, starting with challoc

DevFeed: [Introducing Project Theseus, starting with challoc](<https://devfeed.tech/articles/introducing-project-theseus-starting-with-challoc-133958.md>)

Original publisher: [Read original article](<https://www.chainguard.dev/unchained/introducing-project-theseus-starting-with-challoc>)

Published: 2026-11-13T00:00:00Z

Content type: release

Language: en

Sources: [Chainguard: Unchained](<https://devfeed.tech/sources/chainguard-unchained.md>)

Topics: [Open Source](<https://devfeed.tech/topics/open-source.md>), [Jule](<https://devfeed.tech/topics/jule.md>), [open source maintenance](<https://devfeed.tech/topics/open-source-maintenance.md>), [Linux Kernel](<https://devfeed.tech/topics/linux-kernel.md>), [CI/CD](<https://devfeed.tech/topics/cicd.md>), [Test-driven development](<https://devfeed.tech/topics/tdd.md>)

Tags: [abi](<https://devfeed.tech/tags/abi.md>), [agent](<https://devfeed.tech/tags/agent.md>), [agents](<https://devfeed.tech/tags/agents.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-agents](<https://devfeed.tech/tags/ai-agents.md>), [ai-coding](<https://devfeed.tech/tags/ai-coding.md>), [ai-coding-agents](<https://devfeed.tech/tags/ai-coding-agents.md>), [amd](<https://devfeed.tech/tags/amd.md>), [anthropic](<https://devfeed.tech/tags/anthropic.md>), [apache](<https://devfeed.tech/tags/apache.md>), [chainguard](<https://devfeed.tech/tags/chainguard.md>), [components](<https://devfeed.tech/tags/components.md>), [launches](<https://devfeed.tech/tags/launches.md>), [open-source](<https://devfeed.tech/tags/open-source.md>), [open-source-components](<https://devfeed.tech/tags/open-source-components.md>), [patching](<https://devfeed.tech/tags/patching.md>), [project](<https://devfeed.tech/tags/project.md>), [reactive](<https://devfeed.tech/tags/reactive.md>), [rust](<https://devfeed.tech/tags/rust.md>), [source](<https://devfeed.tech/tags/source.md>)

### AI overview

Chainguard's Project Theseus aims to replace core open-source components to prevent recurring classes of bugs; its first experimental component, challoc, is a Rust allocator designed as a drop-in replacement for glibc malloc, jemalloc, tcmalloc, and mimalloc. It retains the standard C allocator ABI and jemalloc's extended API, allowing existing programs to use it without recompilation. The article says Verus proves 86% of its functions and 94% of its lines, with model checking at system-call and assembly boundaries; isolation, randomized address ranges, and invalid-free checks are intended to make heap bugs harder to exploit. Challoc adapts memory reclamation to cgroup limits and pressure. In cited tests it returned freed memory faster than jemalloc and avoided OOM kills in a workload where jemalloc and mimalloc were killed. Chainguard reports lower memory use and competitive throughput in application benchmarks, including Valkey and Redis, but these are company-run comparisons. Wolfi rebuilds and tests covered several packages on two architectures. The article recommends trying challoc only in test environments; it remains experimental.

### Source excerpt

Chainguard launches Project Theseus to move beyond reactive patching with safer open source components, starting with the Rust-based challoc allocator.

## After the October Surprise: Why Automated Patching for Infrastructure Starts with a System of Record

DevFeed: [After the October Surprise: Why Automated Patching for Infrastructure Starts with a System of Record](<https://devfeed.tech/articles/after-the-october-surprise-why-automated-patching-for-infrastructure-starts-with-a-system-of-record-131139.md>)

Original publisher: [Read original article](<https://netboxlabs.com/blog/automated-infrastructure-patching-system-of-record/>)

Author: Lee Pender

Published: 2026-11-12T19:36:27Z

Content type: article

Language: en

Sources: [NetBox Labs](<https://devfeed.tech/sources/netbox-labs.md>)

Topics: [NetBox](<https://devfeed.tech/topics/netbox.md>), [configuration-management](<https://devfeed.tech/topics/configuration-management.md>)

Tags: [ansible](<https://devfeed.tech/tags/ansible.md>), [automated](<https://devfeed.tech/tags/automated.md>), [blog](<https://devfeed.tech/tags/blog.md>), [infrastructure](<https://devfeed.tech/tags/infrastructure.md>), [netbox](<https://devfeed.tech/tags/netbox.md>), [october](<https://devfeed.tech/tags/october.md>), [patching](<https://devfeed.tech/tags/patching.md>), [system-of-record](<https://devfeed.tech/tags/system-of-record.md>)

### AI overview

The post argues that infrastructure patching speed and safety depend on an accurate system of record: advisories identify vulnerable software, while inventory and context show where it runs, its dependencies and owners, and how to reach it for remediation. Incomplete visibility leaves teams searching spreadsheets and consoles, and unreachable devices fall back to manual work. It proposes measuring automation reach--the share of in-scope infrastructure automation can access and manage--and using exposure and reachability to prioritize response. Its workflow is a continuous loop from advisory through identification, prioritization, bounded remediation, and verification, expanding one change class at a time. It presents NetBox and NetBox Assurance as the inventory baseline, paired with tools such as Red Hat Ansible for targeted jobs. These are the post's claims; the supplied text does not provide independent validation.

### Source excerpt

The post After the October Surprise: Why Automated Patching for Infrastructure Starts with a System of Record appeared first on NetBox Labs.

## Are Causal Neuron Circuits Unique and Transferable?

DevFeed: [Are Causal Neuron Circuits Unique and Transferable?](<https://devfeed.tech/articles/are-causal-neuron-circuits-unique-and-transferable-128768.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.13885>)

Author: \\"Omer Tekin Yavuz (\\.Ileti\\c{s}im Yaz{\\i}l{\\i}m)

Published: 2026-11-11T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [Neural Network](<https://devfeed.tech/topics/neural-network.md>), [causality](<https://devfeed.tech/topics/causality.md>)

Tags: [circuits](<https://devfeed.tech/tags/circuits.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [gemma](<https://devfeed.tech/tags/gemma.md>), [gemma-3](<https://devfeed.tech/tags/gemma-3.md>), [interpretability](<https://devfeed.tech/tags/interpretability.md>), [llama](<https://devfeed.tech/tags/llama.md>), [machine-learning](<https://devfeed.tech/tags/machine-learning.md>), [mlp](<https://devfeed.tech/tags/mlp.md>), [model](<https://devfeed.tech/tags/model.md>), [neuron](<https://devfeed.tech/tags/neuron.md>), [patching](<https://devfeed.tech/tags/patching.md>), [qwen3](<https://devfeed.tech/tags/qwen3.md>)

### AI overview

Testing all 4,096 subsets of selected 12-neuron MLP groups, the authors find that causal effects are often redundant and input-specific. On a Qwen3-0.6B capital task, zero-ablating the selected neurons lowers answer margin by 27%, but replacing their activations with values from other questions barely changes it; the activations are nearly constant, with no evidence they carry question-specific content. On two-hop relation chains, patching selected neurons with first-hop activations shifts answers toward the bridge entity in 9/11 chains for Qwen3-1.7B and Llama-3.2-1B, and 8/10 for Gemma-3-1B; random neurons never do. A few neurons often suffice, and disjoint subsets can independently cause the flip. Effects carry over to rephrasings of the same chain but are much weaker on other tested chains. The study examines neuron-level interventions and alternative subsets, not full circuit connectivity.

### Source excerpt

arXiv:2611.13885v1 Announce Type: new Abstract: Interpretability demonstrations often show that intervening on a group of neurons changes a model's answer. A recent example removes an 18-neuron group in Qwen3-32B and the model answers "Simba" instead of "Scar" to a two-hop Lion King riddle. Such a demonstration establishes that the group matters, but not how much of it is necessary, whether other groups could do the same, or whether the effect carries over to other inputs. We ask: when intervening on a group of neurons changes a model's answer, how much of that group is actually necessary? For every selected group we measure all 4,096 subsets of 12 MLP neurons. On a two-alternative capital task (Qwen3-0.6B), zero-ablating 12 selected neurons lowers the answer margin by 27% while random neurons barely change it, yet replacing the same neurons with their values from other questions also barely changes it: their activations are nearly constant across questions, and we find no evidence that they carry question-specific content. On two-hop relation chains, patching 12 neurons with activations from the first-hop question shifts the answer to the bridge entity (from Scar to Simba): the greedy answer starts with the bridge in 9/11 chains for Qwen3-1.7B, 9/11 for Llama-3.2-1B and 8/10 for Gemma-3-1B; random neurons never do. The full group is rarely necessary: a few neurons often suffice, and in many chains disjoint subsets each flip the answer on their own, in an exploratory analysis most often when the model is less confident. Group effects transfer to rephrasings of the same chain but are much weaker on other tested chains. We study neuron-level causal interventions and alternative subsets; we do not recover the full connectivity of circuits.

## Patch Tuesday - November 2026

DevFeed: [Patch Tuesday - November 2026](<https://devfeed.tech/articles/patch-tuesday-november-2026-128015.md>)

Original publisher: [Read original article](<https://www.rapid7.com/blog/post/em-patch-tuesday-november-2026>)

Author: Adam Barnett

Published: 2026-11-10T21:55:37Z

Content type: news

Language: en

Sources: [Rapid7 Cybersecurity](<https://devfeed.tech/sources/rapid7-cybersecurity-blog.md>)

Topics: [known exploitable vulnerabilities](<https://devfeed.tech/topics/known-exploitable-vulnerabilities.md>), [microsoft 365](<https://devfeed.tech/topics/microsoft-365.md>)

Tags: [2026](<https://devfeed.tech/tags/2026.md>), [classic-outlook](<https://devfeed.tech/tags/classic-outlook.md>), [cve](<https://devfeed.tech/tags/cve.md>), [cve-2026-124194](<https://devfeed.tech/tags/cve-2026-124194.md>), [cybersecurity-company](<https://devfeed.tech/tags/cybersecurity-company.md>), [exploitation](<https://devfeed.tech/tags/exploitation.md>), [exploited](<https://devfeed.tech/tags/exploited.md>), [exposure-assessment-platform](<https://devfeed.tech/tags/exposure-assessment-platform.md>), [exposure-management](<https://devfeed.tech/tags/exposure-management.md>), [file](<https://devfeed.tech/tags/file.md>), [hash](<https://devfeed.tech/tags/hash.md>), [malicious](<https://devfeed.tech/tags/malicious.md>), [managed-detection-and-response](<https://devfeed.tech/tags/managed-detection-and-response.md>), [managed-security-solutions](<https://devfeed.tech/tags/managed-security-solutions.md>), [microsoft](<https://devfeed.tech/tags/microsoft.md>), [november-2026](<https://devfeed.tech/tags/november-2026.md>), [ntlm](<https://devfeed.tech/tags/ntlm.md>), [outlook](<https://devfeed.tech/tags/outlook.md>), [patch-tuesday](<https://devfeed.tech/tags/patch-tuesday.md>), [patching](<https://devfeed.tech/tags/patching.md>), [security](<https://devfeed.tech/tags/security.md>), [spoofing](<https://devfeed.tech/tags/spoofing.md>), [vulnerabilities](<https://devfeed.tech/tags/vulnerabilities.md>), [vulnerability-management](<https://devfeed.tech/tags/vulnerability-management.md>)

### AI overview

The excerpt reports 1,092 Microsoft product vulnerabilities for November 2026 Patch Tuesday, including 1,102 in Windows, and says Microsoft knows of in-the-wild exploitation for one vulnerability. CVE-2026-124194 affects classic Outlook: opening a specially crafted malicious file can silently connect to an attacker-controlled location and leak the user's NTLM v2 hash. The text notes that the large rise in vulnerability volume has not been accompanied by a rise in known in-the-wild exploitation, and questions whether patching more flaws will improve security outcomes given the possibility of new bugs. The excerpt ends mid-sentence, so it does not establish the full affected-version scope or remediation details.

### Source excerpt

Microsoft is publishing 1,092 own-product vulnerabilities on November 2026 Patch Tuesday, including 1,102 vulnerabilities in Windows. Until a few months ago, these volumes would have been unimaginable, but if anything, this month feels like a slight reprieve. The jury is still out on whether patching many more vulnerabilities will lead to better security outcomes; every code change to resolve a known vulnerability could easily introduce new hidden treasure for researchers or threat actors to unearth in the future. Despite the vast uptick in volume during 2026, there has been no accompanying increase in the number of vulnerabilities known to be exploited in the wild. Microsoft is aware of exploitation in the wild for just one of the vulnerabilities published today, with no known public disclosure at all this month. Outlook: zero-day spoofing Imagine coming home and checking your mailbox to find a postcard. You admire the non-specific image of a tropical beach at sunset, and flip it over to scan the handwriting on the back, which looks vaguely familiar. It's addressed to you, but you can't make out the signature, and the message is short and generic. Is that real handwriting, or has this been printed? Probably marketing material. You put the card in the recycling and, since life is busy, you move on and forget about it. Somehow, you remain entirely unaware that, immediately after reading the postcard, you took out your phone and sent high-resolution images of your front door key to a number you've never contacted before. Is this sketch just some rather questionable science fiction? Yes, hopefully, but what if that non-specific postcard was instead a malicious file in your Outlook inbox? Now we're talking about CVE-2026-124194, an exploited-in-the-wild spoofing vulnerability. Merely opening a specially crafted malicious file causes Outlook to silently connect to an attacker-controlled location and leak the user's NTLM v2 hash. All supported versions of classic Outlook

## Seeing Isn't Enough: Unlocking Visual Reasoning by Bridging the Reasoning-Access Gap with TIGER

DevFeed: [Seeing Isn't Enough: Unlocking Visual Reasoning by Bridging the Reasoning-Access Gap with TIGER](<https://devfeed.tech/articles/seeing-isn-t-enough-unlocking-visual-reasoning-by-bridging-the-reasoning-access-gap-with-tiger-125216.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.10552>)

Author: Gregory Kang Ruey Lau (Department of Computer Science, National University of Singapore, Singapore, CNRS@CREATE, Singapore), Minh Huynh Nguyen (Department of Computer Science, National University of S

Published: 2026-11-10T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>), [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [llm-reasoning](<https://devfeed.tech/topics/llm-reasoning.md>), [Artificial Intelligence](<https://devfeed.tech/topics/ai.md>), [grpo](<https://devfeed.tech/topics/grpo.md>), [LLM evaluation / benchmarking](<https://devfeed.tech/topics/llm-evaluation-benchmarking.md>)

Tags: [benchmarks](<https://devfeed.tech/tags/benchmarks.md>), [charts](<https://devfeed.tech/tags/charts.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cv](<https://devfeed.tech/tags/cs-cv.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [dpo](<https://devfeed.tech/tags/dpo.md>), [group-relative-policy-optimization](<https://devfeed.tech/tags/group-relative-policy-optimization.md>), [grpo](<https://devfeed.tech/tags/grpo.md>), [image](<https://devfeed.tech/tags/image.md>), [improvements](<https://devfeed.tech/tags/improvements.md>), [input](<https://devfeed.tech/tags/input.md>), [language-models](<https://devfeed.tech/tags/language-models.md>), [large-language-models](<https://devfeed.tech/tags/large-language-models.md>), [multimodal](<https://devfeed.tech/tags/multimodal.md>), [optimization](<https://devfeed.tech/tags/optimization.md>), [patching](<https://devfeed.tech/tags/patching.md>), [policy](<https://devfeed.tech/tags/policy.md>), [reasoning](<https://devfeed.tech/tags/reasoning.md>), [rlvr](<https://devfeed.tech/tags/rlvr.md>), [sft](<https://devfeed.tech/tags/sft.md>), [spatial](<https://devfeed.tech/tags/spatial.md>), [spatial-reasoning](<https://devfeed.tech/tags/spatial-reasoning.md>), [synthetic](<https://devfeed.tech/tags/synthetic.md>)

### AI overview

The paper argues that multimodal language models can read text in images yet struggle to apply their text-based reasoning to that visual content. TIGER targets this reasoning-access gap by finding synthetic text/image pairs where a model solves the text version but fails on the rendered image, then post-training on the image inputs. The authors report that GRPO, as well as SFT and DPO variants, improves performance on several real-world visual benchmarks despite training only on synthetic rendered-text images without multimodal annotations. TIGER also adds gains when combined with existing RLVR visual-reasoning methods. Activation analyses support the interpretation that training helps connect visual inputs to text reasoning states.

### Source excerpt

arXiv:2611.10552v1 Announce Type: cross Abstract: Multimodal Large Language Models (MLLMs) inherit strong textual reasoning capabilities from their language backbones and text pretraining, but often fail to apply the same level of competency to visual inputs. We show that this degradation cannot be attributed to only perception failures, by evaluating models on semantically-equivalent text and rendered-text image inputs: models can accurately read text in the images and recover near-text performance when provided their own transcriptions, but just cannot reason well on the images directly. We characterize this discrepancy as a modality reasoning-access gap, where models fail to route perceived visual content into existing reasoning machinery accessible with text. To bridge this, we propose TIGER (Text-to-Image Gap-targeted training for Enhanced Reasoning), which first mines synthetic modality counterfactuals where a model succeeds on a text problem but fails when provided the corresponding rendered image, and post-trains the model on the image input via Group Relative Policy Optimization (GRPO), along with SFT and DPO variants. Importantly, although trained exclusively on synthetic rendered-text images with no multimodal annotations, TIGER results in model performance improvements across diverse real-world visual benchmarks including geometric diagrams (MathVerse), multi-disciplinary sciences (EMMA, MMMU), complex charts (ChartQAPro), and embodied spatial reasoning (RoboSpatial-Home). TIGER also complements existing state-of-the-art RLVR visual reasoning methods, yielding further performance gains. Analysis involving reasoning-subspace activation and activation patching provide support indicating that TIGER helps to align visual-to-reasoning pathways toward text reasoning states. Our implementation is available at https://github.com/minhngh/Tiger.

## Breach of Confidence: 09 November 2026

DevFeed: [Breach of Confidence: 09 November 2026](<https://devfeed.tech/articles/breach-of-confidence-09-november-2026-123946.md>)

Original publisher: [Read original article](<https://javvadmalik.com/2026/11/09/breach-of-confidence-09-november-2026/>)

Author: j4vv4d

Published: 2026-11-09T12:20:37Z

Content type: news

Language: en

Sources: [Cyber Newsroom](<https://devfeed.tech/sources/cyber-newsroom.md>)

Topics: [trust](<https://devfeed.tech/topics/trust.md>), [cve remediation](<https://devfeed.tech/topics/cve-remediation.md>), [known exploitable vulnerabilities](<https://devfeed.tech/topics/known-exploitable-vulnerabilities.md>)

Tags: [access-control](<https://devfeed.tech/tags/access-control.md>), [alerts](<https://devfeed.tech/tags/alerts.md>), [android](<https://devfeed.tech/tags/android.md>), [antivirus](<https://devfeed.tech/tags/antivirus.md>), [app](<https://devfeed.tech/tags/app.md>), [blog](<https://devfeed.tech/tags/blog.md>), [botnet](<https://devfeed.tech/tags/botnet.md>), [cve-2026-119977](<https://devfeed.tech/tags/cve-2026-119977.md>), [devices](<https://devfeed.tech/tags/devices.md>), [employee](<https://devfeed.tech/tags/employee.md>), [file](<https://devfeed.tech/tags/file.md>), [google](<https://devfeed.tech/tags/google.md>), [ip](<https://devfeed.tech/tags/ip.md>), [metasploit](<https://devfeed.tech/tags/metasploit.md>), [patching](<https://devfeed.tech/tags/patching.md>), [phishing](<https://devfeed.tech/tags/phishing.md>)

### AI overview

The excerpt reports several security incidents: Google disrupted the XCPE Android botnet, which had about one million devices, and banned the TV box maker, though the devices remain active. Fake nuclear alerts in Ukraine and Russia exploited information the alert app accepted rather than breaking the alert system. Some White House devices had rented out their IP addresses to strangers, with 236 guest check-ins and check-outs. A tobacco company's secret recipes were accessible to all employees on a file share, while Metasploit activity reportedly went undetected by antivirus. The excerpt also flags CVE-2026-119977 as a vulnerability exploited during WordPress updates, when sites are exposed during maintenance. It offers no technical remediation details.

### Source excerpt

Breach of Confidence: 09 November 2026 Someone asked me this week to look at their telly because "it's gone a bit slow". I looked at the GBP 30 box tucked behind it and decided some questions are best left unasked. The Premier League was on, ... Read more Published Date: Nov 09, 2026 (7 hours, 41 minutes ago) Vulnerabilities has been mentioned in this article. CVE-2026-119977

## Dynamic Tokenization via Reinforcement Patching: End-to-end Training and Zero-shot Transfer

DevFeed: [Dynamic Tokenization via Reinforcement Patching: End-to-end Training and Zero-shot Transfer](<https://devfeed.tech/articles/dynamic-tokenization-via-reinforcement-patching-end-to-end-training-and-zero-shot-transfer-123110.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2603.26097>)

Author: Yulun Wu, Sravan Kumar Ankireddy, Samuel Sharpe, Nikita Seleznev, Dehao Yuan, Hyeji Kim, Nam H. Nguyen

Published: 2026-11-09T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>), [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [Artificial Intelligence](<https://devfeed.tech/topics/ai.md>), [Reinforcement learning](<https://devfeed.tech/topics/reinforcement-learning.md>), [Time Series](<https://devfeed.tech/topics/time-series.md>), [Representation learning](<https://devfeed.tech/topics/representation-learning.md>)

Tags: [adaptive](<https://devfeed.tech/tags/adaptive.md>), [compression](<https://devfeed.tech/tags/compression.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [data-driven](<https://devfeed.tech/tags/data-driven.md>), [deep-learning](<https://devfeed.tech/tags/deep-learning.md>), [forecasting](<https://devfeed.tech/tags/forecasting.md>), [patching](<https://devfeed.tech/tags/patching.md>), [reinforcement-learning](<https://devfeed.tech/tags/reinforcement-learning.md>), [sequence](<https://devfeed.tech/tags/sequence.md>), [stat-ml](<https://devfeed.tech/tags/stat-ml.md>), [time-series](<https://devfeed.tech/tags/time-series.md>), [tokenization](<https://devfeed.tech/tags/tokenization.md>), [training](<https://devfeed.tech/tags/training.md>), [zero-shot](<https://devfeed.tech/tags/zero-shot.md>)

### AI overview

Reinforcement Patching (ReinPatch) jointly trains a sequence patching policy and its downstream backbone using reinforcement learning. It treats patch-boundary placement as a discrete decision process optimized with Group Relative Policy Gradient, avoiding continuous relaxations. The method can enforce a desired compression rate and supports hierarchical modeling. On time-series forecasting datasets, the authors report performance compared with state-of-the-art data-driven patching strategies. Its detached design also lets the patching module be extracted as a standalone patcher. The supplied text does not give specific benchmark results or describe zero-shot transfer despite the title.

### Source excerpt

arXiv:2603.26097v2 Announce Type: replace Abstract: Efficiently aggregating spatial or temporal horizons to acquire compact representations has become a unifying principle in modern deep learning models, yet learning data-adaptive representations for long-horizon sequence data, especially continuous sequences like time series, remains an open challenge. While fixed-size patching has improved scalability and performance, discovering variable-sized, data-driven patches end-to-end often forces models to rely on soft discretization, specific backbones, or heuristic rules. In this work, we propose Reinforcement Patching (ReinPatch), a novel framework that jointly optimizes a sequence patching policy and its downstream sequence backbone model with reinforcement learning. By formulating patch boundary placement as a discrete decision process optimized via Group Relative Policy Gradient (GRPG), ReinPatch bypasses the need for continuous relaxations and performs dynamic patching policy optimization in a natural manner. Moreover, our method allows strict enforcement of a desired compression rate, freeing the downstream backbone to scale efficiently, and natively supports multi-level hierarchical modeling. We evaluate ReinPatch on time-series forecasting datasets, where it demonstrates compelling performance compared to state-of-the-art data-driven patching strategies. Furthermore, our detached design allows the patching module to be extracted as a standalone foundation patcher, providing the community with visual and empirical insights into the segmentation behaviors preferred by a purely performance-driven neural patching strategy.

## Patching the Blind Spots: Reversible LLM Rule Overlay for Frozen Fraud Detectors

DevFeed: [Patching the Blind Spots: Reversible LLM Rule Overlay for Frozen Fraud Detectors](<https://devfeed.tech/articles/patching-the-blind-spots-reversible-llm-rule-overlay-for-frozen-fraud-detectors-123298.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.08714>)

Author: Bo Qu, Zhurong Wang, Frank Zahradnik, Ming Ouyang

Published: 2026-11-09T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>), [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [online-fraud-prevention](<https://devfeed.tech/topics/online-fraud-prevention.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>)

Tags: [benchmark](<https://devfeed.tech/tags/benchmark.md>), [bias](<https://devfeed.tech/tags/bias.md>), [classification](<https://devfeed.tech/tags/classification.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [e-commerce](<https://devfeed.tech/tags/e-commerce.md>), [fraud](<https://devfeed.tech/tags/fraud.md>), [gradient-boosted-trees](<https://devfeed.tech/tags/gradient-boosted-trees.md>), [llm](<https://devfeed.tech/tags/llm.md>), [machine-learning](<https://devfeed.tech/tags/machine-learning.md>), [overlay](<https://devfeed.tech/tags/overlay.md>), [patching](<https://devfeed.tech/tags/patching.md>)

### AI overview

The paper proposes patching a frozen fraud detector with a reversible rule overlay instead of retraining it. LLMs generate candidate rules from features the model underweights according to SHAP; empirical screening and calibration turn supported rules into additive log-lift adjustments at the production cut-off. Because single-LLM gains varied and reversed across datasets, the authors average logits from three diverse backbones, trading peak gains for a bounded worst case. On a selection-biased e-commerce detector, continued training regressed, while the overlay averaged +2.26% recall at equal precision and +1.60% precision at equal recall across cut-offs spanning 30-57% precision and 23-63% recall. On seven designed public benchmark variants, it raised precision at equal recall on each and kept recall at equal precision near neutral. Amendments are readable predicates identifying the fraud patterns they cover.

### Source excerpt

arXiv:2611.08714v1 Announce Type: new Abstract: In industrial tabular fraud classification, gradient-boosted trees remain state-of-the-art, but even a well-trained model has blind spots, and retraining the deployed model is not always safe. We propose a framework that patches such a model without modifying it: LLMs propose rules anchored on features the model underweights by SHAP, and the resulting noisy pool is screened for empirical support and calibrated into an additive, reversible log-lift overlay that runs at the production cut-off. Single-LLM benefit is high-variance and inverts between datasets, so we pool three diverse backbones by equal-weight logit averaging, trading peak gain for a bounded worst case. On a real-world e-commerce fraud detector trained on selection-biased data, continued training regresses even when carefully tuned, whereas the overlay gains on average +2.26% recall at equal precision ($\Delta$R@P) and +1.60% precision at equal recall ($\Delta$P@R) over production cut-offs spanning 30-57% precision and 23-63% recall. On a public benchmark of seven designed bias-shift variants it raises $\Delta$P@R on every variant and keeps $\Delta$R@P near neutral. Each amendment is a readable predicate naming the fraud pattern it covers.

## Scratchpad Patching: Decoupling Compute from Patch Size in Byte-Level Language Models

DevFeed: [Scratchpad Patching: Decoupling Compute from Patch Size in Byte-Level Language Models](<https://devfeed.tech/articles/scratchpad-patching-decoupling-compute-from-patch-size-in-byte-level-language-models-123141.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2605.09630>)

Author: Lin Zheng, Vasilisa Bashlovkina, Timothy Dozat, Dan Garrette, Laura Rimell, Joshua Maynez

Published: 2026-11-09T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [Chinchilla scaling law](<https://devfeed.tech/topics/chinchilla-scaling-law.md>), [LLM evaluation / benchmarking](<https://devfeed.tech/topics/llm-evaluation-benchmarking.md>)

Tags: [cache](<https://devfeed.tech/tags/cache.md>), [code](<https://devfeed.tech/tags/code.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [efficiency](<https://devfeed.tech/tags/efficiency.md>), [entropy](<https://devfeed.tech/tags/entropy.md>), [inference](<https://devfeed.tech/tags/inference.md>), [lag](<https://devfeed.tech/tags/lag.md>), [language-models](<https://devfeed.tech/tags/language-models.md>), [patching](<https://devfeed.tech/tags/patching.md>), [quality](<https://devfeed.tech/tags/quality.md>), [size](<https://devfeed.tech/tags/size.md>)

### AI overview

Scratchpad Patching (SP) addresses patch lag in byte-level language models: predictions within a patch otherwise rely on a stale representation until that patch is fully observed, and larger patches widen this lag. SP inserts transient scratchpads to aggregate observed bytes and refresh patch-level context. It triggers them using next-byte prediction entropy, allocating compute selectively to information-dense regions and allowing inference-time compute to be adjusted after training. In experiments on natural language and code, SP improved quality at the same patch size. At 16 bytes per patch, SP models matched or closely approached the byte-level baseline on downstream evaluations, with a 16x smaller KV-cache over patches and roughly 3-4x less inference compute than the baseline.

### Source excerpt

arXiv:2605.09630v2 Announce Type: replace-cross Abstract: Tokenizer-free language models eliminate the tokenizer step of the language modeling pipeline by operating directly on bytes; patch-based variants further aggregate contiguous byte spans into patches for efficiency. However, the average patch size chosen at the model design stage governs a tight trade-off: larger patches reduce compute and KV-cache footprint, but degrade modeling quality. We trace this trade-off to patch lag: until a patch is fully observed, byte predictions within it must rely on a stale representation from the previous patch to preserve causality; this lag widens as patches grow larger. We introduce Scratchpad Patching (SP), which inserts transient scratchpads inside each patch to aggregate the bytes seen so far and refresh patch-level context for subsequent predictions. SP triggers scratchpads using next-byte prediction entropy, selectively allocating compute to information-dense regions and enabling post-hoc adjustment of inference-time compute. Across experiments on natural language and code, SP improves model quality at the same patch size; for example, even at $16$ bytes per patch, SP-augmented models match or closely approach the byte-level baseline on downstream evaluations while using a $16\times$ smaller KV-cache over patches and requiring roughly $3$-$4\times$ less inference compute than the byte-level baseline.

## Google Cloud Adds Self-Managed VMware Cloud Foundation on Bare Metal

DevFeed: [Google Cloud Adds Self-Managed VMware Cloud Foundation on Bare Metal](<https://devfeed.tech/articles/google-cloud-adds-self-managed-vmware-cloud-foundation-on-bare-metal-122449.md>)

Original publisher: [Read original article](<https://www.infoq.com/news/2026/11/google-cloud-self-managed-vmware/>)

Author: Steef-Jan Wiggers

Published: 2026-11-08T07:16:00Z

Content type: news

Language: en

Sources: [InfoQ](<https://devfeed.tech/sources/infoq.md>)

Topics: [VMware Cloud Foundation](<https://devfeed.tech/topics/vcf-9-1.md>), [Google Cloud Platform (GCP)](<https://devfeed.tech/topics/google-cloud.md>), [private cloud](<https://devfeed.tech/topics/private-cloud.md>), [vcf networking](<https://devfeed.tech/topics/vcf-networking.md>), [Cloud Networking](<https://devfeed.tech/topics/cloud-networking.md>)

Tags: [22](<https://devfeed.tech/tags/22.md>), [architecture-design](<https://devfeed.tech/tags/architecture-design.md>), [aws](<https://devfeed.tech/tags/aws.md>), [bare-metal](<https://devfeed.tech/tags/bare-metal.md>), [broadcom](<https://devfeed.tech/tags/broadcom.md>), [cloud](<https://devfeed.tech/tags/cloud.md>), [cloud-computing](<https://devfeed.tech/tags/cloud-computing.md>), [customer](<https://devfeed.tech/tags/customer.md>), [development](<https://devfeed.tech/tags/development.md>), [devops](<https://devfeed.tech/tags/devops.md>), [esxi](<https://devfeed.tech/tags/esxi.md>), [google](<https://devfeed.tech/tags/google.md>), [google-cloud](<https://devfeed.tech/tags/google-cloud.md>), [google-cloud-self-managed-vmware](<https://devfeed.tech/tags/google-cloud-self-managed-vmware.md>), [hardware](<https://devfeed.tech/tags/hardware.md>), [infrastructure](<https://devfeed.tech/tags/infrastructure.md>), [licensing](<https://devfeed.tech/tags/licensing.md>), [lifecycle](<https://devfeed.tech/tags/lifecycle.md>), [managed](<https://devfeed.tech/tags/managed.md>), [news](<https://devfeed.tech/tags/news.md>), [nsx](<https://devfeed.tech/tags/nsx.md>), [oracle](<https://devfeed.tech/tags/oracle.md>), [patching](<https://devfeed.tech/tags/patching.md>), [regions](<https://devfeed.tech/tags/regions.md>), [root](<https://devfeed.tech/tags/root.md>), [solution](<https://devfeed.tech/tags/solution.md>), [vmware](<https://devfeed.tech/tags/vmware.md>), [vmware-cloud-foundation](<https://devfeed.tech/tags/vmware-cloud-foundation.md>)

### AI overview

Google Cloud's self-managed VMware option runs VCF 9.x on dedicated Z3 bare metal in 22 regions, with root access to ESXi, vCenter, and NSX. Google provisions the environment, then customers or partners handle upgrades, patching, and portable VCF subscriptions. This contrasts with VMware Engine, which runs VCF 5.2, limits vCenter access to role-based privileges, and leaves lifecycle management to Google. Setup uses a console or Terraform toolkit; prerequisites include VLAN-based networking, an 8896-byte VPC MTU, and management, vSAN, vMotion, and two NSX subnets. Storage uses local NVMe vSAN, with NetApp Volumes available as an external datastore. The customer-controlled model offers access to the current VCF major version but returns operational and licensing responsibilities to the customer; named partners can provide managed operations. Google gives no pricing, SLA, regional list, or migration path from VMware Engine. Its cited TCO savings lack a disclosed baseline or methodology, and the quoted customer describes an early assessment rather than production use.

### Source excerpt

Google Cloud has made a self-managed VMware solution generally available, running VMware Cloud Foundation 9.x on dedicated Z3 bare metal with root access across ESXi, vCenter and NSX in 22 regions. It sits beside the Google-managed VMware Engine and matches customer-operated offerings from AWS and Oracle, shifting lifecycle, patching and Broadcom licensing back to the customer. By Steef-Jan Wiggers

## Why are we all worked up about atomicity when applying hot-patches if we can just suspend all the threads?

DevFeed: [Why are we all worked up about atomicity when applying hot-patches if we can just suspend all the threads?](<https://devfeed.tech/articles/why-are-we-all-worked-up-about-atomicity-when-applying-hot-patches-if-we-can-just-suspend-all-the-threads-122103.md>)

Original publisher: [Read original article](<https://devblogs.microsoft.com/oldnewthing/20261106-00/?p=112838/>)

Author: Raymond Chen

Published: 2026-11-06T15:00:00Z

Content type: article

Language: en

Sources: [Raymond Chen](<https://devfeed.tech/sources/raymond-chen.md>)

Topics: [patches](<https://devfeed.tech/topics/patches.md>), [synchronization](<https://devfeed.tech/topics/synchronization.md>)

Tags: [atomic](<https://devfeed.tech/tags/atomic.md>), [code](<https://devfeed.tech/tags/code.md>), [function](<https://devfeed.tech/tags/function.md>), [old-new-thing](<https://devfeed.tech/tags/old-new-thing.md>), [patches](<https://devfeed.tech/tags/patches.md>), [patching](<https://devfeed.tech/tags/patching.md>), [process](<https://devfeed.tech/tags/process.md>), [processors](<https://devfeed.tech/tags/processors.md>), [suspend](<https://devfeed.tech/tags/suspend.md>), [threads](<https://devfeed.tech/tags/threads.md>), [time](<https://devfeed.tech/tags/time.md>), [update](<https://devfeed.tech/tags/update.md>), [variable](<https://devfeed.tech/tags/variable.md>), [windows](<https://devfeed.tech/tags/windows.md>), [windows-update](<https://devfeed.tech/tags/windows-update.md>), [x86](<https://devfeed.tech/tags/x86.md>)

### AI overview

Atomic activation matters because suspending a process's existing threads does not prevent a new, runnable thread from being created with CreateRemoteThread while patching is underway. On processors with variable-length instructions such as x86, a suspended thread may also be poised to execute bytes that become the tail of a replacement instruction; the post says the original instruction must be at least as large as the replacement to avoid that case. Atomicity still cannot guard against concurrent, non-atomic WriteProcessMemory changes, which the post describes as already broken and outside supported hot-patching use.

### Source excerpt

Suspending all the threads stops the threads that are there now, those are not the only possible sources of change. The post Why are we all worked up about atomicity when applying hot-patches if we can just suspend all the threads? appeared first on The Old New Thing.

[Next page](<https://devfeed.tech/tags/patching.md?cursor=WyIyMDI2LTExLTA2VDE1OjAwOjAwKzAwOjAwIiwgIjNkZmU5M2FhLWMyMWUtNGMyMi05NDZhLTViMDgwMDkxODM2YSJd>)