ai-agent-security
Audited by Socket on Sep 19, 2026
3 alerts found:
Securityx2AnomalySUSPICIOUS: the skill is internally consistent with its stated purpose, but that purpose is to equip an AI agent with offensive security capabilities against other agent systems. It contains explicit credential-theft, exfiltration, command-execution, persistence, stealth, and autonomous-action playbooks, making it a high-risk exploit skill rather than a benign workflow aid.
The artifact is an explicitly labeled AI-agent red-team payload catalog with defensive-validation material. It contains multiple deliberately malicious example programs and payloads, including credential harvesting, command execution, reverse shells, persistence, destructive actions, phishing, vector-store poisoning, and data exfiltration. There is no evidence in the supplied fragment of automatic malware execution or hidden package behavior; the primary risk arises if the documented examples are run or incorporated into production systems. The evil_mcp_server.py and rugpull_server.py examples should only be used in isolated, authorized labs and must not be deployed as ordinary MCP servers.
The fragment is a red-team test-plan document, not executable malware. It does not itself perform the described actions, and the stated scope and cleanup controls are consistent with authorized testing. However, it contains highly actionable credential-exfiltration, sandbox-escape, persistence, SSRF, and destructive-agent procedures. Treat it as sensitive dual-use security content and verify that it is intentionally included, access-controlled, and not accompanied elsewhere by executable payloads or install-time hooks.