INSIGHTS
200+ vulnerabilities open-weight AI was able to exploit
Every assessment Anyday has run, in aggregate: how often AI gets from a public advisory to a working PoC, how far it gets when it fails, and what the attempt costs.
Chrome, Linux kernel, Firefox and WindowsOpen-weight modelsUpdated November 2026
No public exploit
80%
of the vulnerabilities AI built a working exploit for had no public exploit to copy.
EPSS
91%
score below 1%. EPSS predicts exploitation in the wild, not whether an exploit can be built.
CVSS
6%
are rated Critical. Severity does not find them.
The signals teams prioritise with
EPSS score
- Below 1%91%
- 1–10%5%
- 10–50%3%
- 50%+1%
- Unknown2%
CVSS severity
- Low1%
- Medium17%
- High75%
- Critical6%
MEDIAN COST OF A WORKING EXPLOIT
$25.42
CHEAPEST
$6.36
How to read this
Every assessment was run by open-weight models from the public advisory and patch alone. Each CVE page names the model behind its own result. A result counts as a working exploit only when the run ends in control of the target, graded by the same ladder each CVE page uses. A published proof of concept that only reproduces the bug is not counted as a public exploit.