Skip to content

Latest commit

 

History

History
226 lines (181 loc) · 11.5 KB

File metadata and controls

226 lines (181 loc) · 11.5 KB

How agent-audit-kit compares

Context for decision-makers evaluating MCP / AI-agent static scanners in April 2026.

Verifiable claims only. If you find a claim here that no longer holds, file an issue and we will correct it (best-effort — no fixed clock).

At a glance

agent-audit-kit Microsoft AGT Snyk Agent Scan (ex Invariant) Semgrep Multimodal SAST Lakera Guard
License MIT MIT Proprietary Proprietary + OSS rules Proprietary
Scope Static scanner + compliance evidence Runtime governance (policy engine + mesh) Static + runtime (post-acquisition) Multimodal SAST Runtime guardrail
Account / cloud required No No (but Azure-native paths) Yes Optional Yes
Cloud round-trip No No Yes (findings leave your repo) Optional Yes
Compliance-evidence PDF Yes (EU AI Act, SOC 2, ISO 27001+42001, HIPAA, NIST AI RMF, Singapore, India DPDP, Alabama DPPA, Tennessee SB 1580) No (runtime policies, no audit PDFs) No (findings only) No No
Regional / US-state compliance Yes (India DPDP, Singapore, Alabama, Tennessee) No No No No
Signed rule bundle Yes (Sigstore) Partial (SLSA provenance on releases) No (proprietary) No No
Deterministic (reproducible CI) Yes Yes (sub-ms policy enforcement) No (multi-model analysis) Partial No
Public CVE-to-rule ledger Yes (CHANGELOG.cves.md) No (internal cadence) No No No
MCP Security Index / leaderboard Yes (weekly, 500+ servers) No No No No
Pin + drift verification of tool surface Yes Yes (via Agent Runtime rings) No No No
OWASP Agentic Top 10 coverage 10/10 10/10 Partial Partial Partial

Microsoft Agent Governance Toolkit (Apr 2 2026)

Microsoft open-sourced AGT under MIT license on April 2 2026, with broad coverage on Apr 16 (Help Net Security, InfoWorld). It is the first major-vendor entry to claim 10/10 OWASP Agentic Top 10 coverage with deterministic sub-millisecond policy enforcement.

AGT is an ally in positioning, not a head-on competitor. It validates the category we've been shipping for months; different tool, different layer:

  • agent-audit-kit runs at CI time and ship time. It's a static scanner + compliance evidence generator. You run it on a repo to catch issues before deployment and to produce auditor-ready PDFs (EU AI Act Art. 15, ISO 42001, Alabama DPPA, Tennessee SB 1580, etc.).
  • Microsoft AGT runs at runtime. Agent OS (policy engine), Agent Mesh (A2A comms), Agent Runtime (dynamic execution rings), Agent SRE (reliability), Agent Compliance (automated evidence collection), Agent Marketplace, Agent Lightning (RL training governance). It enforces policies as agents execute.

Use both: agent-audit-kit tells an auditor your design is sound; Microsoft AGT tells them your runtime actually behaved. Differentiation is the compliance-evidence PDF stack (EU/US state-by-state), the public CVE-to-rule ledger, the signed rule bundle, and the MCP Security Index leaderboard — none of which are AGT goals.

Microsoft AGT ships Python / TypeScript / Rust / Go / .NET. Integrations already operational in Dify, LlamaIndex, OpenAI Agents SDK, Haystack, LangGraph, PydanticAI. agent-audit-kit integrates with AGT findings on the roadmap (a future --runtime-policies microsoft-agt flag can cross-check our static rules against a deployed AGT policy set).

SnapLogic AI Gateway + Trusted Agent Identity (Apr 16 2026)

SnapLogic announced enterprise iPaaS primitives for agent identity + governance. Signal: "agent identity" is now a vendor-category, not a feature. agent-audit-kit's pin + verify commands already cover the "is this the agent I expected?" question for MCP tool surface; SnapLogic extends the same idea to cross-enterprise RPA / iPaaS flows. Complementary, not competing.

What we are NOT better at

  • Multi-model analysis. Snyk's acquisition of Invariant Labs bought them a proprietary corpus + a multi-model pipeline. Their ToxicSkills recall numbers (claimed 90–100%) are out of reach for a deterministic scanner. If you need semantic coverage on skills you don't author, you want both: Snyk for that and agent-audit-kit for compliance evidence + pin/verify + a public CVE-to-rule ledger.

  • Hosted dashboards. We ship SARIF for GitHub Security tab. If you want a hosted triage dashboard, a commercial product is easier.

  • Vulnerability research. We ship detection for disclosed MCP CVEs on a best-effort basis, tracked in a public ledger. We do NOT originate CVE research — that's Invariant, Palo Alto Unit 42, HiddenLayer, Check Point, etc.

When to pick agent-audit-kit

  • You need an auditor-ready compliance report (EU AI Act high-risk obligations Aug 2 2026).
  • Your environment is air-gapped (defense, finance, healthcare) and no data can leave the repo.
  • You need reproducible CI — the same scan on the same commit must produce byte-identical output.
  • You have regional compliance obligations in India (DPDP Act) or Singapore (Agentic AI Governance Framework).
  • You want to pin + verify your MCP tool surface over time.
  • You need a public scanner whose rule set your security team can read, audit, and fork.

When to pick a commercial scanner instead

  • You have zero in-house security review capacity and want a vendor SLA, phone number, and on-call.
  • You need a hosted triage dashboard beyond SARIF → GitHub.
  • You prefer to pay for multi-model semantic analysis of skills/tools from third parties.

The honest state of the market

The OSS agent-security-scanner category is crowded (May 2025–Apr 2026): snyk/agent-scan, cisco-ai-defense/mcp-scanner, riseandignite/mcp-shield, mcpshield/mcpshield, affaan-m/agentshield, HeadyZhang/agent-audit, plus Semgrep's Multimodal SAST.

What's empty in the category: compliance evidence mapped to specific regulatory articles; deterministic reproducibility; a public CVE-to-rule ledger; a pinning + drift primitive; a public leaderboard. agent-audit-kit occupies that space.

Last reviewed: 2026-04-18.


Feature matrix vs the OSS scanners

Merged here from docs/comparison.md in v0.3.86. Three comparison pages coexisted, the README linked only this one, and the other two drifted unread.

AgentAuditKit vs Competitors

Feature AgentAuditKit mcp-scan Snyk Agent Agent Audit Microsoft AGT
Rules 320 ~10 ~15 57 N/A (runtime)
MCP config scanning Yes No Yes No No
Hook injection detection Yes No No No No
Trust boundary analysis Yes No No No Yes
Secret exposure scanning Yes No Yes No No
Supply chain analysis Yes No Yes No No
Agent instruction scanning Yes No No No No
Tool poisoning detection Yes Yes Yes No No
Tool pinning / rug pull Yes Yes No No No
Taint analysis (@tool) Yes No No Yes No
A2A protocol scanning Yes No No No No
Multi-agent discovery Yes No Yes No No
OWASP Agentic Top 10 10/10 0/10 Partial 10/10 10/10
OWASP MCP Top 10 10/10 Partial Partial 0/10 0/10
Compliance frameworks 12 0 0 0 3
SARIF output Yes No Yes No No
Auto-fix mode Yes No No No No
Security scoring Yes No No No No
Pre-commit hook Yes No No No No
GitHub Action Yes No Yes No No
Runtime proxy Yes No No No Yes
Offline / no network Yes No No Yes Yes
Zero dependencies Yes* No No No No

*Only click + pyyaml required.

When to Use Each

  • AgentAuditKit: Comprehensive static + config scanning, compliance reporting, CI/CD integration
  • mcp-scan: Quick tool description poisoning check via cloud API
  • Snyk Agent Scan: Enterprise multi-agent MDM with cloud backend
  • Agent Audit: Academic-quality taint analysis for LangChain/CrewAI code
  • Microsoft AGT: Runtime policy enforcement with execution rings

These tools are complementary. Use AgentAuditKit alongside runtime tools for defense-in-depth.


GitLab Agentic SAST 18.11 (2026-04-17)

Merged here from docs/comparison-gitlab-agentic-sast.md in v0.3.86.

GitLab 18.11 (2026-04-17) shipped Agentic SAST behind the Ultimate tier. This is a no-marketing, dated-source comparison so AAK consumers can pick the right tool for their stack.

Dimension AgentAuditKit GitLab Agentic SAST 18.11
License MIT, OSS Proprietary, Ultimate-tier paywall
Distribution PyPI + Marketplace + Docker + VS Code ext GitLab CI / Premium offering only
Rule count 320 Not publicly disclosed
OWASP Agentic Top 10 mapping Per-rule, public JSON manifest Claimed; mapping not published
MCP Top 10 mapping Per-rule Not advertised
AICM (CSA) mapping Yes Not advertised
Out-of-band corpus refresh aak corpus update (signed) No equivalent — rules ship on product release cadence
SARIF diff (regression-only gating) aak diff --baseline ... --current ... Not advertised
VS Code extension Yes (open-source) No
PR-title indirect prompt injection AAK-PRTITLE-IPI-001 (CVSS 9.4) Not advertised in 18.11 changelog
MCP function-hijacking detection AAK-MCP-FHI-001 (arXiv 2604.20994) Not advertised
Atlassian MCP CVEs AAK-MCP-ATLASSIAN-CVE-2026-27825/27826 Generic SAST may catch; no per-CVE rule
Sigstore-attested releases Yes Not applicable
Self-scan / dogfood gate .github/workflows/self-scan.yml (PR-blocking) Not applicable to a managed product

Where GitLab is genuinely stronger

  • DAST + IAST integration: GitLab integrates SAST + DAST + IAST in one pipeline. AAK is SAST + supply-chain only.
  • Multi-tenant org-level dashboards: GitLab's Vulnerability Reports aggregate across the org. AAK ships SARIF; consumers wire their own dashboard.
  • Ecosystem coverage: GitLab's traditional SAST has rule packs for ~30 languages. AAK targets the agent / MCP slice specifically.

Where AAK is genuinely stronger

  • Same-day defense for new payload families: Out-of-band signed corpus refresh decouples threat-data updates from product releases.
  • PR-title IPI rule: First-to-market on Comment-and-Control class (CVSS 9.4, 2026-04-25 disclosure).
  • MCP function-hijacking rule: First-to-market on the BFCL FHI class (arXiv 2604.20994, 2026-04-23, 70-100% ASR).
  • Free + OSS: No tier paywall, no per-seat licensing.
  • SARIF regression gating: aak diff lets PR-blocking workflows gate on newly_introduced only — eliminates the "huge backlog blocks every PR" failure mode that GitLab Ultimate users report.

Sources

Last updated: 2026-04-27