ml-supply-chain
Safeguard articles tagged "ml-supply-chain" — guides, analysis, and best practices for software supply chain and application security.
17 articles
Understanding model poisoning and backdoored model weights
A poisoned model looks like any other checkpoint file. Here's how model poisoning attacks work, real incidents on Hugging Face, and how detection and provenance checks catch them.
Analysis of pickle file deserialization vulnerabilities i...
CVE-2025-32434 shows PyTorch's "safe" weights_only loading could still be bypassed for code execution — a pickle deserialization vulnerability with real supply-chain consequences.
Understanding membership inference attacks against traine...
Membership inference attacks let adversaries confirm if your data trained a model, exposing privacy leakage in ML models and training data inference risks.
The torchtriton Dependency Confusion Attack on PyTorch-Ni...
How a namespace gap on PyPI let a malicious "torchtriton" package hijack PyTorch-nightly installs for five days, and what it teaches about ML supply chain security.
vLLM CVE-2025-62164: Tensor Deserialization RCE
vLLM 0.10.2-0.11.0 deserialized user-supplied PyTorch tensors via torch.load() in the Completions API. Memory corruption, potential RCE.
vLLM CVE-2025-66448: Auto-Map RCE via Model Configs
A critical RCE in vLLM allows malicious model configs to bypass trust_remote_code=False. We analyze the bug, the patch, and what every vLLM operator should do.
ShadowMQ: 30+ RCE Flaws Across AI Inference Engines
Oligo Security disclosed ShadowMQ in November 2025: ZeroMQ-and-pickle deserialization patterns copy-pasted across vLLM, Meta Llama, TensorRT-LLM, and others. We dissect the pattern.
Hugging Face as Malware CDN and Exfiltration Backend: The DPRK-Linked npm Campaign of May 2026
OX Security disclosed a DPRK-aligned campaign that abused Hugging Face as a malware host and data-exfiltration backend, using public repos to serve second-stage payloads and private datasets to receive stolen developer secrets.
The Fake OpenAI 'privacy-filter' Model: How a Typosquat Hit #1 on Hugging Face in May 2026
A repository named Open-OSS/privacy-filter impersonated OpenAI's release, copied its model card verbatim, and shipped a loader.py that pulled an infostealer. It reached #1 trending with ~244,000 downloads before removal.
PyTorch CVE-2025-32434: weights_only=True No Longer Safe
A critical PyTorch RCE bypassed the safety property of torch.load(weights_only=True). We analyze the bug and explain why safetensors should now be the default.
OpenSSF Model Signing v1.0: Sigstore for ML
OpenSSF launched Model Signing v1.0 in April 2025 with Sigstore integration. NVIDIA NGC adopted it the same month. We explain what it signs, how to verify, and where the gaps are.
nullifAI: Broken Pickles and the Hugging Face Detection Gap
ReversingLabs disclosed two malicious Hugging Face models that evaded Picklescan by using broken 7z-packed PyTorch archives. We unpack the technique.
Self-healing security runs on Safeguard.
Your first fix PR is minutes away.
No sales call required, even your agent can complete the purchase over MCP.