ai-threat-testing

SUSPICIOUS: purpose and content are internally aligned, but the skill is an offensive security framework for AI agents and materially increases attack capability against real targets. No clear credential theft, covert exfiltration, or malicious installer is present, so this is high-risk offensive functionality rather than confirmed malware.

The fragment is defensive supply-chain security documentation containing substantial dual-use and explicitly offensive MLflow and MCP exploitation guidance. It includes descriptions of pickle deserialization RCE, artifact manipulation, default-credential testing, persistence, callbacks, and arbitrary command submission, but no executable implementation or actual malicious behavior is present in the supplied text. Obfuscation is absent. The principal risk is misuse of the documented procedures rather than package malware.

The content documents a potentially harmful adversarial-input technique that could bypass ML-based authorization when applied to a vulnerable endpoint. It contains no evidence of malware or unauthorized system activity in the supplied fragment. Its primary risk is dual-use attack guidance and the underlying insecure practice of using a single classifier output as an authorization decision.