OpenAI Tests GPT-6 Astra on 20 Recent Chrome V8 Vulnerabilities
OpenAI has put GPT-6 Astra through ExploitBench, a new cybersecurity benchmark built from 20 high-severity Chrome V8 vulnerabilities disclosed across 13 stable Chrome releases between June and August 2026.
The benchmark tests whether an AI agent can achieve arbitrary code execution in V8 and official Chrome releases for Linux. It is narrower than a general cyber benchmark and more useful for exactly that reason: the task is not writing plausible vulnerability analysis. It is getting exploit development to work against recent browser targets.
Why This Benchmark Matters
Most public cybersecurity evaluations are easy to contaminate. Old CVEs, public proof-of-concept writeups, and heavily discussed exploit chains can leak into model training data or agent memory. ExploitBench is built from a recent three-month window, which makes memorisation less useful and forces the model to perform more of the exploit-development work at evaluation time.
The evaluation also avoids the neatness of a simple 100% ceiling. Some included vulnerabilities may not permit arbitrary code execution under the benchmark constraints, so a perfect score may not be achievable. That matters because it keeps the task closer to real security work, where a bug can be real but still resist a working exploit in a constrained environment.
Astra Versus Sol
Astra achieved substantially higher arbitrary code-execution rates than GPT-5.6 Sol on ExploitBench while using far fewer output tokens. The public benchmark material frames the result as a capability and efficiency gap rather than a single headline pass rate.
That distinction is important. A model that can search exploit space with fewer generated tokens is not just cheaper to run. It can attempt more branches, recover from failed exploit paths faster, and fit longer tool-use loops into the same budget. In cybersecurity, token efficiency turns into exploration budget.
The Safety Reading
ExploitBench lands at the centre of Astra’s rollout problem. Astra is already being treated as a model with unusually strong cybersecurity capability, with early access routed first to approved defenders. A benchmark showing stronger exploit success against recent V8 targets supports that caution.
The test does not prove broad autonomous cyber capability by itself. Browser V8 exploitation is a specialised domain. But it is a domain with real operational stakes, and arbitrary code execution against recent Chrome releases is not a toy task.
Key Numbers
| Metric | Value |
|---|---|
| Vulnerabilities | 20 |
| Severity | High |
| Target | Chrome V8 |
| Chrome releases | 13 stable releases |
| Vulnerability window | June to August 2026 |
| Execution target | Arbitrary code execution |
| Comparator | GPT-5.6 Sol |
The practical read is simple: Astra is not merely better at talking about cybersecurity. It is being measured on whether it can turn recent browser vulnerabilities into working exploit chains, and it is outperforming OpenAI’s previous frontier model on that axis.