Brief, mostly facts

Endor Labs benchmark finds agents pass security tests far less often

Across 200 tasks, Claude Code with Claude Fable 5.1 passes 87.2 percent of functional tests but only 37.4 percent of security tests. The benchmark covers 108 Python projects and 77 CWEs.

Read the source at endorlabs.com

What the research found

Endor Labs Agent Security League benchmark (200 tasks, 108 Python projects, 77 CWEs) shows top agents pass functional tests often but pass security tests far less, with Claude Code plus Claude Fable 5.1 at 87.2% functional and 37.4% secure.

Claims and how well the evidence supports them

Claude Code with Claude Fable 5.1 scores 87.2% functional and 37.4% secure 98%
Benchmark covers 200 tasks from 108 open-source Python projects across 77 CWE classes 99%
Secure pass rates across listed agents range from about 4% to 37% 94%

Evidence

endorlabs.com https://endorlabs.com/research/ai-code-security-benchmark

Actionability rated medium.