ExploitGym Hardened (stronger exploit mitigations) (ExploitGym Hardened)
We show this table for reference; we do not rank on it.
A hardened variant of ExploitGym that adds address-space randomization and stack canaries to userspace targets, enables the V8 sandbox and uses stricter kernel configurations where supported.
Benchmark score on ExploitGym Hardened — November 11, 2026
We compile the ExploitGym Hardened rows from benchmark-owner or independent runs and provider self-reports. GPT-6 Cyber leads the table at 36.9%, followed by GPT-6 Astra (29.7%) and Claude Mythos Preview (5.0%). We do not use these results to rank models overall.
GPT-6 Cyber
OpenAI
GPT-6 Astra
OpenAI
Claude Mythos Preview
Anthropic
5 modelsAgenticCurrentDisplay onlyUpdated November 11, 2026
Benchmark score table (5 models)
| Rank | Model / configuration | Score | Parameters (B) | Open / closed |
|---|---|---|---|---|
| 1 | GPT-6 CyberOpenAI | 36.9% | Not reported | Closed |
| 2 | GPT-6 AstraOpenAI | 29.7% | Not reported | Closed |
| 3 | Claude Mythos PreviewAnthropic | 5.0% | Not reported | Closed |
| 4 | GPT-5.5OpenAI | 2.3% | Not reported | Closed |
| 5 | GPT-5.4OpenAI | 0.3% | Not reported | Closed |
How to read this leaderboard
Operator receipt: 5 sourced rows are currently displayable on this page; the leading published row is GPT-6 Cyber at 36.9%.
Honest limit: Provider-run variant; per-task mitigations and the exact task subset are not published.
Among the reported ExploitGym Hardened rows, GPT-6 Cyber is first at 36.9%. The third row is 31.9 points behind. The broader top-10 range is 36.6 points, so the table still separates the published systems.
5 models have been evaluated on ExploitGym Hardened. The benchmark falls in the Agentic category. We keep these results separate from the weighted global scoring system, so they remain source-specific evidence. ExploitGym Hardened is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.
About ExploitGym Hardened
Year
2026
Tasks
ExploitGym tasks with stronger exploit mitigations
Format
Intended-exploit success rate
Difficulty
Exploitation against mitigations
OpenAI reports the intended-exploit success rate, the same metric as standard ExploitGym: GPT-6 Cyber 36.94% and GPT-6 Astra 29.69%, against 48.45% and 42.43% on the standard suite. The mitigation modes come from the ExploitGym authors (UC Berkeley RDI): the benchmark repository ships hardened userspace builds, a V8 sandbox build and kernel defense settings, and the paper's mitigation table reports successes with defenses on for Claude Mythos Preview (45 of 898 tasks), GPT-5.5 (21) and GPT-5.4 (3), with Claude Opus 4.6, Claude Opus 4.7, Gemini 3.1 Pro and GLM-5.1 at zero. BenchLM stores paper rows as a share of the 898 tasks and OpenAI's rows as reported on its 869-task v1 implementation, so the two are not a like-for-like comparison. Display-only and separate from the ExploitGym lane.
Freshness and provenance
Version
ExploitGym Hardened (stronger exploit mitigations)
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 ExploitGym Hardened measure?
A hardened variant of ExploitGym that adds address-space randomization and stack canaries to userspace targets, enables the V8 sandbox and uses stricter kernel configurations where supported.
Which model scores highest on ExploitGym Hardened?
GPT-6 Cyber by OpenAI currently leads with a score of 36.9% on ExploitGym Hardened.
How many models are evaluated on ExploitGym Hardened?
5 AI models have published results on ExploitGym Hardened in the BenchLM catalog.
Compare top models on ExploitGym Hardened
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 7,000+ subscribers.
One email each week. Unsubscribe anytime.