Research · Scanned April 2026 · Inkog v1.1.0

State of AI Agent Security 2026

We ran Inkog over 561 open-source AI-agent repositories and recorded everything it reported. These are automated, unreviewed results at the most sensitive setting. Read them as a map of where agent code tends to go wrong, not as a count of confirmed vulnerabilities.

561

open-source agent repositories scanned (577 attempted, 16 failed to scan)

11,705

findings in total, about 21 per repository

85.6%

had at least one finding of any severity

62.6%

had at least one finding rated critical or high

Most findings are risks, not exploits

Inkog sorts every finding into a tier. Fewer than 7% were exploitable vulnerabilities. The bulk were structural: an agent that can do something risky with nothing in the code to limit or approve it.

Exploitable vulnerability · a concrete path an attacker could use
789 (6.7%)
Structural risk pattern · a missing limit, check or approver
8,492 (72.6%)
Hardening · best practice, low urgency
2,039 (17.4%)
Unclassified
385 (3.3%)

What came up most

  • Unbounded loops5,397
  • Missing audit logging1,371
  • Over-reliance on model output941
  • Missing rate limits938
  • Token bombing610
  • eval / exec on dynamic input573
  • Hardcoded credentials (4 main types)416
  • Missing human oversight335
  • SQL injection via model output149
  • Sensitive data in logs139
  • SSRF123
  • Unsafe deserialisation117

Treat the loop count with the most caution: loop detection is the check most prone to false positives, and none of these results were reviewed. We left the numbers as the scanner produced them rather than adjust them by hand.

EU AI Act

143 repositories (25%) had at least one finding mapped to Article 14, human oversight: an agent that can take a consequential action with no approval step in the code. Article 14 only binds high-risk systems, and most of these repositories are not one, so this measures readiness rather than non-compliance.

Method

  • 40 GitHub search queries for agent frameworks and MCP servers, top 100 results per query by stars, de-duplicated, 20+ stars, no forks.
  • Each repository shallow-cloned on 4 April 2026 and scanned with Inkog v1.1.0 on the comprehensive policy: every detector, no confidence filter.
  • No findings were reviewed by hand for this report. The deeper, reviewed work on a subset is in the study below.

Read next: What 561 repositories taught us about AI agent security