agentshield

# AgentShield — AI Agent Security Auditor ## What it Does Cherry-picked from `everything-claude-code` (affaan-m/everything-claude-code v1.9.0) Security-first auditing for AI agent code, skills, and prompts. ## When to Use - Before deploying a new agent or skill to AI OS Corp - When reviewing third-party skill code - As part of GATE_SECURITY in plugin-integration workflow - Regular audits of existing agent prompts ## Audit Checklist ### Prompt Security ``` □ No prompt injection vulnerabilities (user input directly in system prompt) □ No role-confusion patterns ("ignore previous instructions") □ No data exfiltration paths (agent sending data to external URLs without auth) □ No hardcoded credentials or API keys in prompts □ Scope is clearly bounded (agent can't exceed its defined role) ``` ### Code Security ``` □ No exec() or eval() on user input □ No subprocess with shell=True on untrusted input □ File access scoped to AI OS root only □ No unrestricted network calls □ Dependencies locked (pinned versions) ``` ### Data Handling ``` □ No PII logging □ No sensitive data in plain-text files □ Memory (Mem0) stores only non-sensitive summaries □ Outputs sanitized before display ``` ## Quick Audit Command ```bash # Run AgentShield on a skill or agent file agentshield audit ./skills/my-skill/SKILL.md agentshield audit ./corp/agents/my-agent.md agentshield scan-dir ./skills/ # Scan all skills ``` ## Security Score Output ``` AgentShield Score: 87/100 ✅ No prompt injection patterns ✅ No hardcoded secrets ⚠️ WARN: Subprocess call without validation (line 42) ✅ Data handling: SAFE Decision: CONDITIONAL — fix warning before deploy ``` ## Notes - Source: github.com/affaan-m/everything-claude-code (cherry-pick) - 113 contributors | v1.9.0 (Mar 2026) - Owner: Dept 10 (Security GRC) — run before every new skill - Works alongside: trivy (infrastructure), strix-scan (full pipeline) - Full documentation available in `security-guide.md` within this folder.