V8 Unique Confirmed Security Issues
Safety · 2026-07-21
Google’s V8 vulnerability-hunting evaluation counts unique confirmed security issues found without safety guardrails under a fixed but undisclosed number of model invocations. Because the invocation budget and evaluated target corpus are not published, Atlas retains this count as a diagnostic rather than a normalized capability score.
Top models (higher is better)
| Model | Score |
|---|---|
| Gemini 3.5 Flash Cyber | 55.0 |
| Gemini 3.5 Flash | 47.0 |
| Opus 4.6 | 36.0 |