ai-security

SUSPICIOUS/HIGH-RISK skill: internally coherent as an AI security red-team skill, but it gives an AI agent offensive security capabilities against external AI systems and local MCP configurations. Install sources appear to be mainstream registries and same-org tooling rather than clear malware, so this is better classified as a vulnerable/high-risk offensive skill than confirmed malicious code.

The code contains deliberate retrieval-poisoning functionality and a limited embedding-inversion demonstration. It writes attacker-controlled payloads to local files and can optimize those documents for high query similarity. It does not show malware indicators such as data exfiltration, reverse shells, credential theft, persistence, cryptomining, or destructive actions. File paths and glob patterns are unrestricted by the code, so use with untrusted arguments can cause unintended local file reads or writes. The primary risk is misuse against RAG or vector-search systems, not host compromise.

The fragment is an attack-research implementation and documentation for optimizing adversarial language-model suffixes and discussing model extraction. It does not itself show conventional npm/package malware or system compromise, but it has meaningful dual-use security risk, and the OPSEC section includes advice that could help evade abuse detection. Assessment is limited to the visible portion.