OPSWAT adds LLM scanning to MetaDefender Aether v3.3.0

OPSWAT has released MetaDefender Aether v3.3.0, an update that adds LLM scanning to analyze threats inside AI model files and risks that surface while a model responds, according to datanet.co.kr. InSecCyberSecurity, the company's Korean partner, announced the release on October 20.

The LLM scanning is added to the product's existing five-stage zero-day threat detection system. Static analysis covers 7 model file formats: Pickle, PyTorch, TensorFlow, Keras, SafeTensors, GGUF and ONNX. To counter detection evasion, the file type is determined by actual byte data rather than by the file extension.

On the dynamic side, MetaDefender Aether sends adversarial prompts to models that expose an OpenAI-compatible Chat Completions endpoint. The analysis draws on more than 50 attack techniques and over 1000 adversarial prompts spread across 9 attack categories.

Results from both the static and dynamic analysis are mapped to MITRE ATLAS, the framework that catalogues adversarial machine learning attack techniques.

Earlier testing context

In September 2026, AMTSO updated its agentic security product testing guidelines and sandbox evaluation framework with AI features, adding an "LLM-as-Sample Dynamic Analysis" metric. In a 2026 sandbox evaluation by Venak Security, MetaDefender Aether Adaptive Sandbox detected 19 of 20 AI-generated malware samples, the highest result among 4 sandboxes evaluated.