What happened
On Oct 12, OPSWAT released MetaDefender Aether v3.3.0 with LLM Scanning: Layer-2 static analysis parses 7 LLM file formats (Pickle, PyTorch, TensorFlow, Keras, SafeTensors, GGUF, ONNX) without executing them, surfacing embedded payloads via 30+ threat indicators; Layer-3 dynamic analysis runs live model endpoints through 50+ adversarial techniques and 1,000+ prompts across nine attack categories, mapped to MITRE ATLAS. OPSWAT positions this among the first sandbox vendors to add LLM scanning, aligned with AMTSO's new 'LLM-as-Sample Dynamic Analysis' KPI.
Why it matters
Model files are now a documented attack surface (CVE-2026-80047 in Hugging Face Transformers, NemoClaw/Ollama takeover, PuzzleMask). OPSWAT is extending established five-layer zero-day detection (an AMTSO-tested top-performer on AI malware) into model-file vetting for an adoption flow (Hugging Face pulls into prod) that most security tooling ignores — a practical, deployable answer to the open-weight/third-party model risk gap highlighted by the Microsoft-NVIDIA whitepaper in the feed.
Applicability
Teams pulling third-party/open-weight models into production, MLOps platforms, or air-gapped deployments should evaluate MetaDefender Aether LLM Scanning for pre-deployment model vetting; sandbox/scanning vendors should track AMTSO's LLM-as-Sample KPI.