# safetensors

Published articles for safetensors.

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

## In a pickle: Why every pickle defense in machine learning is bypassed the same way

DevFeed: [In a pickle: Why every pickle defense in machine learning is bypassed the same way](<https://devfeed.tech/articles/in-a-pickle-why-every-pickle-defense-in-machine-learning-is-bypassed-the-same-way-102875.md>)

Original publisher: [Read original article](<https://webflow.sysdig.com/blog/in-a-pickle-why-every-pickle-defense-in-machine-learning-is-bypassed-the-same-way>)

Author: Lydia Graslie

Published: 2026-10-27T00:00:00Z

Content type: article

Language: en

Sources: [Sysdig](<https://devfeed.tech/sources/sysdig-blog.md>)

Topics: [Adversarial attacks](<https://devfeed.tech/topics/adversarial-attacks.md>), [Machine learning](<https://devfeed.tech/topics/machine-learning.md>), [Keras](<https://devfeed.tech/topics/keras.md>), [model-deployment](<https://devfeed.tech/topics/model-deployment.md>), [MLOps](<https://devfeed.tech/topics/mlops.md>), [hugging face](<https://devfeed.tech/topics/hugging-face.md>), [CI/CD](<https://devfeed.tech/topics/cicd.md>), [scikit-learn](<https://devfeed.tech/topics/scikit-learn.md>)

Tags: [ai-supply-chain-security](<https://devfeed.tech/tags/ai-supply-chain-security.md>), [best-practices](<https://devfeed.tech/tags/best-practices.md>), [blast-radius](<https://devfeed.tech/tags/blast-radius.md>), [bypass](<https://devfeed.tech/tags/bypass.md>), [ci](<https://devfeed.tech/tags/ci.md>), [code-execution](<https://devfeed.tech/tags/code-execution.md>), [cve-2025-32434](<https://devfeed.tech/tags/cve-2025-32434.md>), [cve-2026-24747](<https://devfeed.tech/tags/cve-2026-24747.md>), [falco](<https://devfeed.tech/tags/falco.md>), [keras](<https://devfeed.tech/tags/keras.md>), [keras-safe-mode-bypass](<https://devfeed.tech/tags/keras-safe-mode-bypass.md>), [machine-learning](<https://devfeed.tech/tags/machine-learning.md>), [malicious-pickle-files](<https://devfeed.tech/tags/malicious-pickle-files.md>), [ml-model-security](<https://devfeed.tech/tags/ml-model-security.md>), [nullifai](<https://devfeed.tech/tags/nullifai.md>), [pickle-deserialization](<https://devfeed.tech/tags/pickle-deserialization.md>), [picklescan](<https://devfeed.tech/tags/picklescan.md>), [pytorch](<https://devfeed.tech/tags/pytorch.md>), [pytorch-weights-only-bypass](<https://devfeed.tech/tags/pytorch-weights-only-bypass.md>), [runtime-detection](<https://devfeed.tech/tags/runtime-detection.md>), [safetensors](<https://devfeed.tech/tags/safetensors.md>), [sysdig-secure](<https://devfeed.tech/tags/sysdig-secure.md>)

### AI overview

Python pickle files are executable programs: loading one can run code through its reconstruction instructions, even if the file is malformed after the payload. The excerpt describes six documented bypasses across PyTorch restricted loading, Keras safe_mode, and picklescan; the common weakness is that scanners and loaders interpret the same bytes differently. It identifies safetensors as structurally unable to express a callable, while GGUF and ONNX have no known code execution path but are not presented as proven safe. For models that still require pickle, the recommended layers are current loader versions (PyTorch 2.10.0+; reject legacy Keras .h5; keep NumPy allow_pickle=False), CI scanning, provenance checks, sandboxed loading with restricted egress and no valuable credentials, and runtime monitoring. Runtime monitoring observes downstream effects rather than deserialization itself: the test found all five gadget families triggered Sysdig Secure and four triggered Falco 0.44.1 defaults, with the tested rules detecting activity such as file reads or outbound connections, not the pickle load.

### Source excerpt

Six documented bypasses beat PyTorch weights_only, Keras safe_mode, and picklescan. See which ML formats are safe and what runtime detection catches.

## DreamBooth on Runpod: How to Train for Great Results

DevFeed: [DreamBooth on Runpod: How to Train for Great Results](<https://devfeed.tech/articles/dreambooth-on-runpod-how-to-train-for-great-results-94556.md>)

Original publisher: [Read original article](<https://www.runpod.io/blog/dreambooth-training-runpod-guide>)

Author: Zhen Lu

Published: 2026-10-22T02:42:59Z

Content type: tutorial

Language: en

Sources: [Runpod Blog.](<https://devfeed.tech/sources/runpod-blog.md>)

Topics: [Image](<https://devfeed.tech/topics/image.md>), [bestpractice](<https://devfeed.tech/topics/bestpractice.md>), [GPU](<https://devfeed.tech/topics/gpu.md>)

Tags: [amazing](<https://devfeed.tech/tags/amazing.md>), [automatic1111](<https://devfeed.tech/tags/automatic1111.md>), [best-practices](<https://devfeed.tech/tags/best-practices.md>), [bug](<https://devfeed.tech/tags/bug.md>), [cuda](<https://devfeed.tech/tags/cuda.md>), [disk-space](<https://devfeed.tech/tags/disk-space.md>), [experiments](<https://devfeed.tech/tags/experiments.md>), [great](<https://devfeed.tech/tags/great.md>), [guide](<https://devfeed.tech/tags/guide.md>), [how-to](<https://devfeed.tech/tags/how-to.md>), [images](<https://devfeed.tech/tags/images.md>), [pytorch](<https://devfeed.tech/tags/pytorch.md>), [resume](<https://devfeed.tech/tags/resume.md>), [safetensors](<https://devfeed.tech/tags/safetensors.md>), [setup](<https://devfeed.tech/tags/setup.md>), [train](<https://devfeed.tech/tags/train.md>), [training](<https://devfeed.tech/tags/training.md>), [ubuntu](<https://devfeed.tech/tags/ubuntu.md>), [unique](<https://devfeed.tech/tags/unique.md>), [update](<https://devfeed.tech/tags/update.md>)

### AI overview

Zhen's guide uses a series of Runpod experiments to help users get started with DreamBooth training and troubleshoot common workflow issues. The excerpt lists setup and installation, adding disk space, an xformers error, resuming failed training, reviewing samples, comparing checkpoints with a consistent seed, downloading selected checkpoints with runpodctl, using the web terminal when Jupyter is unavailable, and locating downloaded safetensors files. Specific settings and experiment results are not included in the supplied excerpt.

### Source excerpt

DreamBooth can generate amazing, highly personalized images, but only if you train it well. In this post, Zhen walks through best practices for getting.

## Train Qwen3 on AWS Trainium with torchtitan

DevFeed: [Train Qwen3 on AWS Trainium with torchtitan](<https://devfeed.tech/articles/train-qwen3-on-aws-trainium-with-torchtitan-93901.md>)

Original publisher: [Read original article](<https://aws.amazon.com/blogs/machine-learning/train-qwen3-on-aws-trainium-with-torchtitan/>)

Author: Josh Longenecker

Published: 2026-10-21T20:00:22Z

Content type: tutorial

Language: en

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

Topics: [AWS Trainium](<https://devfeed.tech/topics/aws-trainium.md>), [neuron](<https://devfeed.tech/topics/neuron.md>), [Mixture of Experts (MoE)](<https://devfeed.tech/topics/mixture-of-experts-moe.md>), [AI Infrastructure](<https://devfeed.tech/topics/ai-infrastructure.md>), [AWS AI chips](<https://devfeed.tech/topics/aws-ai-chips.md>), [hugging face](<https://devfeed.tech/topics/hugging-face.md>)

Tags: [advanced-300](<https://devfeed.tech/tags/advanced-300.md>), [all](<https://devfeed.tech/tags/all.md>), [aws](<https://devfeed.tech/tags/aws.md>), [aws-trainium](<https://devfeed.tech/tags/aws-trainium.md>), [backend](<https://devfeed.tech/tags/backend.md>), [config](<https://devfeed.tech/tags/config.md>), [experts](<https://devfeed.tech/tags/experts.md>), [hugging-face](<https://devfeed.tech/tags/hugging-face.md>), [kernels](<https://devfeed.tech/tags/kernels.md>), [loading](<https://devfeed.tech/tags/loading.md>), [mixture-of-experts](<https://devfeed.tech/tags/mixture-of-experts.md>), [modeling](<https://devfeed.tech/tags/modeling.md>), [native](<https://devfeed.tech/tags/native.md>), [qwen3](<https://devfeed.tech/tags/qwen3.md>), [safetensors](<https://devfeed.tech/tags/safetensors.md>), [structure](<https://devfeed.tech/tags/structure.md>), [technical-how-to](<https://devfeed.tech/tags/technical-how-to.md>), [train](<https://devfeed.tech/tags/train.md>), [trainium](<https://devfeed.tech/tags/trainium.md>), [transformers](<https://devfeed.tech/tags/transformers.md>), [v5](<https://devfeed.tech/tags/v5.md>), [with](<https://devfeed.tech/tags/with.md>)

### AI overview

On a single AWS Trainium trn2.48xlarge, torchtitan can train Qwen3 through three paths using the same torchrun launcher and config structure: Transformers v5 via AutoModelForCausalLM, torchtitan's native Qwen3 model selected with config.model_spec, or NKI kernels for mixture-of-experts layers. The Transformers path can load published Hugging Face weights and save standard safetensors; the article reports a first-step loss of 6.54 from published weights versus 12.43 from random initialization for Qwen3-4B-Instruct-2507. The NKI example uses 64 data-parallel ranks without expert parallelism, and reduces Qwen3-30B-A3B to four layers and 64 experts to fit the instance. Custom NKI operators can be attached through torchtitan's override registry, but the rotary-kernel example demonstrates registration and autograd integration, not a performance gain. Pin the walkthrough's versions because the Transformers integration and Trainium overrides can change between torchtitan releases.

### Source excerpt

Train Qwen3 on AWS Trainium with torchtitan three ways on a single trn2.48xlarge: the Hugging Face Transformers v5 modeling backend, torchtitan's native Qwen3, and NKI mixture-of-experts kernels. The same torchrun launcher and config structure work unchanged across all three paths, including loading and saving published Hugging Face safetensors.

## JADEPUFFER evolves: The agentic threat actor deploys ransomware built to destroy AI models

DevFeed: [JADEPUFFER evolves: The agentic threat actor deploys ransomware built to destroy AI models](<https://devfeed.tech/articles/jadepuffer-evolves-the-agentic-threat-actor-deploys-ransomware-built-to-destroy-ai-models-53240.md>)

Original publisher: [Read original article](<https://webflow.sysdig.com/blog/jadepuffer-evolves-the-agentic-threat-actor-deploys-ransomware-built-to-destroy-ai-models>)

Author: Michael Clark

Published: 2026-07-20T00:00:00Z

Content type: article

Language: en

Sources: [Sysdig](<https://devfeed.tech/sources/sysdig-blog.md>)

Topics: [ransomware](<https://devfeed.tech/topics/ransomware.md>), [Threat Research](<https://devfeed.tech/topics/threat-research.md>), [AI Infrastructure](<https://devfeed.tech/topics/ai-infrastructure.md>), [AI Models](<https://devfeed.tech/topics/ai-models.md>), [Cybercrime](<https://devfeed.tech/topics/cybercrime.md>)

Tags: [agentic-ransomware](<https://devfeed.tech/tags/agentic-ransomware.md>), [agentic-threat-actor](<https://devfeed.tech/tags/agentic-threat-actor.md>), [ai-infrastructure](<https://devfeed.tech/tags/ai-infrastructure.md>), [ai-infrastructure-security](<https://devfeed.tech/tags/ai-infrastructure-security.md>), [ai-model-recovery](<https://devfeed.tech/tags/ai-model-recovery.md>), [ai-models](<https://devfeed.tech/tags/ai-models.md>), [ai-pipeline-security](<https://devfeed.tech/tags/ai-pipeline-security.md>), [ai-ransomware](<https://devfeed.tech/tags/ai-ransomware.md>), [ata](<https://devfeed.tech/tags/ata.md>), [cloud-attack](<https://devfeed.tech/tags/cloud-attack.md>), [cloud-threat-detection](<https://devfeed.tech/tags/cloud-threat-detection.md>), [container-escape](<https://devfeed.tech/tags/container-escape.md>), [cve-2025-3248](<https://devfeed.tech/tags/cve-2025-3248.md>), [docker-socket-abuse](<https://devfeed.tech/tags/docker-socket-abuse.md>), [encforge](<https://devfeed.tech/tags/encforge.md>), [exploited](<https://devfeed.tech/tags/exploited.md>), [gguf](<https://devfeed.tech/tags/gguf.md>), [go](<https://devfeed.tech/tags/go.md>), [jadepuffer](<https://devfeed.tech/tags/jadepuffer.md>), [langflow](<https://devfeed.tech/tags/langflow.md>), [llm-security](<https://devfeed.tech/tags/llm-security.md>), [machine-learning-security](<https://devfeed.tech/tags/machine-learning-security.md>), [ml-ransomware](<https://devfeed.tech/tags/ml-ransomware.md>), [model-artifact-encryption](<https://devfeed.tech/tags/model-artifact-encryption.md>), [mysql](<https://devfeed.tech/tags/mysql.md>), [privileged-container](<https://devfeed.tech/tags/privileged-container.md>), [pytorch](<https://devfeed.tech/tags/pytorch.md>), [ransomware](<https://devfeed.tech/tags/ransomware.md>), [ransomware-detection](<https://devfeed.tech/tags/ransomware-detection.md>), [runtime-security](<https://devfeed.tech/tags/runtime-security.md>), [safetensors](<https://devfeed.tech/tags/safetensors.md>), [sysdig-trt](<https://devfeed.tech/tags/sysdig-trt.md>), [threat-research](<https://devfeed.tech/tags/threat-research.md>), [vector-database-security](<https://devfeed.tech/tags/vector-database-security.md>), [yara](<https://devfeed.tech/tags/yara.md>)

### AI overview

Sysdig Threat Research reports that JADEPUFFER, an autonomous threat actor, upgraded its attack on Langflow-based AI infrastructure by deploying ENCFORGE, a Go ransomware payload designed to encrypt model checkpoints, vector databases, training datasets, and embedding indices.

### Source excerpt

JADEPUFFER, the agentic threat actor documented by the Sysdig Threat Research Team, is now using ransomware to destroy trained AI models.