ExploitBench — Internal Port (June–August 2026) (ExploitBench Internal Port)
We show this table for reference; we do not rank on it.
OpenAI's internal port of ExploitBench restricted to V8 and Chrome vulnerabilities disclosed in June to August 2026, to limit contamination from historical exploits.
Benchmark score on ExploitBench Internal Port — November 5, 2026
We compile the ExploitBench Internal Port rows from provider self-reports. GPT-6 Cyber leads the table at 42.0%, followed by GPT-6 Astra (39.0%) and GPT-6.1 Terra (24.5%). We do not use these results to rank models overall.
GPT-6 Cyber
OpenAI
GPT-6 Astra
OpenAI
GPT-6.1 Terra
OpenAI
7 modelsAgenticCurrentDisplay onlyUpdated November 5, 2026
Benchmark score table (7 models)
| Rank | Model / configuration | Score | Parameters (B) | Open / closed |
|---|---|---|---|---|
| 1 | GPT-6 CyberOpenAI | 42.0% | Not reported | Closed |
| 2 | GPT-6 AstraOpenAI | 39.0% | Not reported | Closed |
| 3 | GPT-6.1 TerraOpenAI | 24.5% | Not reported | Closed |
| 4 | GPT-6.1 SolOpenAI | 21.5% | Not reported | Closed |
| 5 | GPT-5.6 SolOpenAI | 5.5% | Not reported | Closed |
| 6 | GPT-6 SolOpenAI | 5.5% | Not reported | Closed |
| 7 | GPT-6 LunaOpenAI | 0.0% | Not reported | Closed |
How to read this leaderboard
Operator receipt: 7 sourced rows are currently displayable on this page; the leading published row is GPT-6 Cyber at 42.0%.
Honest limit: Provider-run port with an unpublished task list; not comparable to the public ExploitBench Cap% score.
Among the reported ExploitBench Internal Port rows, GPT-6 Cyber is first at 42.0%. The third row is 17.5 points behind. The broader top-10 range is 42.0 points, so the table still separates the published systems.
7 models have been evaluated on ExploitBench Internal Port. The benchmark falls in the Agentic category. We keep these results separate from the weighted global scoring system, so they remain source-specific evidence. ExploitBench Internal Port is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.
About ExploitBench Internal Port
Year
2026
Tasks
Recently disclosed V8 and Chrome vulnerabilities (June–August 2026)
Format
Mean arbitrary-code-execution success rate
Difficulty
Browser exploitation
The public ExploitBench uses vulnerabilities going back to 2023 and OpenAI models now score at or near 100% on it, which OpenAI says may be inflated by contamination. The internal port reports mean arbitrary-code-execution success across recently disclosed V8 and Chrome targets: GPT-6 Cyber 42.0%, GPT-6 Astra 39.0%, GPT-5.6 Sol 5.5% (charted as GPT-5.6 under Daybreak Blue). OpenAI's later system cards report the port on a series where GPT-6 Astra scores 31.5%: GPT-6.1 Terra 24.5%, GPT-6.1 Sol 21.5%, GPT-6 Sol 5.5%, GPT-6 Luna 0% and GPT-5.6 Sol 3.5%. BenchLM stores those rows as published, so compare rows from different OpenAI documents with care. OpenAI-run, so display-only and kept separate from the public ExploitBench leaderboard.
Freshness and provenance
Version
ExploitBench — Internal Port (June–August 2026)
Refresh cadence
Static
Staleness state
Current
Question availability
Private provider task set with published aggregate results
BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.
Questions
What does ExploitBench Internal Port measure?
OpenAI's internal port of ExploitBench restricted to V8 and Chrome vulnerabilities disclosed in June to August 2026, to limit contamination from historical exploits.
Which model scores highest on ExploitBench Internal Port?
GPT-6 Cyber by OpenAI currently leads with a score of 42.0% on ExploitBench Internal Port.
How many models are evaluated on ExploitBench Internal Port?
7 AI models have published results on ExploitBench Internal Port in the BenchLM catalog.
Compare top models on ExploitBench Internal Port
Know when it’s worth switching models
The model to choose, the cheaper alternative, and the release we would wait on.
Read a sample issueJoin 6,100+ subscribers.
One email each week. Unsubscribe anytime.