RESEARCHMETAL MANTRA RESEARCH
What the code of AI agents says.
Dated studies built from the public registry: what we scanned, what showed up, our read of it, and what we got wrong along the way.
We scanned the code of 94 of GitHub’s hottest AI agents
Median score 78.5/100; 19 graded CCC; only 13 of the 88 owners are GitHub-verified organizations. Model calls without a fallback, dynamic code execution and vulnerable dependencies were the most common findings.