Skip to main content

Open source agent security

agent-audit-kit

SAST scanner for AI agents. Full OWASP Agentic Top-10 + MCP Top-10.

SAST-style security scanner for agentic AI systems. Catches prompt-injection surfaces, tool-call abuse patterns, MCP STDIO injection (CVE-2026-30623), excessive agency, audit-trail gaps, and the rest of the OWASP Agentic + MCP threat surface, all caught at static-analysis time, before the agent reaches production.

12stars0forks148rulesv0.3.58released 2026-07-23

Quickstart

Drop into any Python project. Outputs SARIF that GitHub Code Scanning understands natively. About 30 seconds.

bash
pip install agent-audit-kit
agent-audit-kit scan ./your-agent-project --format sarif > agent-audit.sarif

Or use the GitHub Action for CI-time scanning on every PR.

Rule catalogue

148 rules across these families. Each rule ships with a working detector, a remediation snippet, and a SARIF severity.

AAK-001..010

OWASP Agentic Top-10

Full 10/10 coverage. Prompt injection, sensitive info disclosure, supply-chain, data + model poisoning, improper output handling, excessive agency, tool integration risks, weak validation, audit trail gaps.

AAK-MCP-001..010

OWASP MCP Top-10

Full 10/10 coverage. STDIO config injection, MCP auth bypass, tool definition tampering, transport confusion, sandbox escape patterns, and the rest of the MCP threat surface.

AAK-CMPL-*

Compliance rule families

Each maps a subset of the rule catalogue.

EU AI ActSOC 2HIPAANIST AI RMFISO 42001India DPDPAlabama HB 351Tennessee SB 1580

What it catches that nothing else does

MCP STDIO config injection

Rule family for CVE-2026-30623: unsanitised MCP STDIO configs that amount to RCE-by-design. Detected at config parse time. Most agent SAST tools don't cover the MCP transport at all.

Tool-call allow-list drift

Flags when an agent gains tool access that isn't in the declared allow-list. Catches scope-creep before a deploy compounds it.

Audit-trail-gap detection

Static-analysis pass that flags agent runs that can't be reconstructed from logs alone. The audit-trail-first principle, enforced at CI time.

Capability-lease enforcement

Flags long-lived keys passed to tool calls when capability-leases (short-lived, scoped, revocable) are the safer pattern. Cross-checks against Okta NHI / Cisco AWI integration points.

Interoperates with

Microsoft Agent Governance Toolkit

OWASP-Agentic 10/10 coverage cross-mapped.

GitHub Marketplace

Listed as a SAST action. Drop into any GitHub Actions workflow.

Use it in your CI today

MIT-licensed. Public roadmap. 48-hour CVE-to-rule SLA. Star the repo to track new rule families as they ship.

Want it customised for your stack? Agent Security Audit: 2 weeks, $4,950. Custom rule pack + audit-trail review + prioritized playbook.