DEV Community

Anders
Anders

Posted on • Originally published at nerq.ai

AI Agent Vulnerability Report — Q1 2026

AI Agent Vulnerability Report — Q1 2026

Published: 2026-03-13 | Author: Nerq Research

Nerq's security crawler scanned 204,000+ AI agents across GitHub, npm, PyPI, HuggingFace, and MCP registries for known vulnerabilities. This is what we found.

Executive Summary

Metric Count
Agents scanned 204,000+
With known CVEs 4,692 (2.3%)
CRITICAL severity 9 agents
HIGH severity 31 agents
Unpatched (no fix available) 12 agents

The Paradox: Popular Agents Are More Likely to Have CVEs

This seems counterintuitive, but the most-used agents are more likely to have known CVEs — not because they're less secure, but because they get more scrutiny. Agents with zero CVE data haven't been audited; that doesn't mean they're safe.

The distinction matters for trust scoring. Nerq's Trust Score v2 handles this with a nuanced Security dimension:

  • CVE found + patched = small penalty, credit for responsiveness
  • CVE found + unpatched = major penalty
  • No CVE data = neutral (score 70, not 100)

Critical Vulnerabilities to Watch

We identified 9 agents with CRITICAL severity CVEs (CVSS >= 9.0). These are agents where:

  • Remote code execution is possible
  • Authentication bypass exists
  • Data exfiltration is trivial

We're not naming specific agents here to allow maintainers time to patch. If you want to check a specific agent:

curl "https://nerq.ai/v1/preflight?target=AGENT_NAME"
Enter fullscreen mode Exit fullscreen mode

The response includes cve_count and the Security component score.

License Classification Results

We also classified licenses for 2,704 agents:

Category Count %
PERMISSIVE (MIT, Apache, BSD) 1,891 70%
COPYLEFT (GPL, AGPL) 389 14%
UNKNOWN 298 11%
PROPRIETARY 126 5%

The 11% with UNKNOWN licenses is a compliance risk — especially for enterprise deployments where license obligations matter.

Recommendations

For agent users:

  1. Run preflight checks before adopting any agent: pip install nerq
  2. Prefer agents with PERMISSIVE licenses for commercial use
  3. Check CVE status regularly — vulnerabilities are disclosed daily

For agent maintainers:

  1. Enable GitHub Security Advisories on your repository
  2. Add a clear license file (SPDX format preferred)
  3. Add a Nerq Trust Badge to your README: [![Nerq Trust](https://nerq.ai/badge/YOUR_AGENT)](https://nerq.ai/safe/YOUR_AGENT)

For platform builders:

  1. Integrate preflight checks into your agent loading pipeline
  2. Use the batch endpoint for checking all agents at startup
  3. Set trust score thresholds appropriate to your risk tolerance

Data Sources

  • CVE data: GitHub Advisory Database (cross-referenced with NVD)
  • License data: SPDX classification from repository license files
  • Download data: npm Registry API, PyPI JSON API
  • Agent index: Nerq Trust Index (204K+ agents, updated continuously)

Full data available via the Nerq API. Free, no auth required. API docs | Python SDK

Top comments (0)