What does ExploitGym measure?
ExploitGym tests whether an agent can develop a working exploit for a specified, real-world vulnerability. Its targets include userspace programs, Google's V8 engine, and the Linux kernel. Success requires obtaining the target's flag through code execution and verifying that the exploit used the intended vulnerability.
The benchmark's current public release contains 869 instances. The provider reports below do not all identify their task revision or subset, so that public-release count does not establish the denominator of an individual result.
OpenAI release results
OpenAI's November 5, 2026 launch report gives GPT-6 Cyber a 48.5% success rate, compared with 32.1% for GPT-5.6 Cyber. These are the launch report's rounded values. Both models have restricted cybersecurity access; the availability links identify their access terms.
| Model configuration | Score | Harness and setup | Source and date |
|---|---|---|---|
| GPT-6 Cyber (Effort unspecified)Approved Daybreak Red organizations | 48.5% | Harness and effort not specified | OpenAI, 2026-11-05; Provider-reported launch values; restricted cyber models |
| GPT-5.6 Cyber (Effort unspecified)Approved Daybreak Red organizations | 32.1% | Harness and effort not specified | OpenAI, 2026-11-05; Provider-reported launch values; restricted cyber models |
OpenAI system-card results
The GPT-6 Cyber system card, section 7.0.5, reports 48.45% for GPT-6 Cyber and 42.43% for GPT-6 Astra. These values retain the card's precision and comparison cohort. The card does not identify each row's reasoning effort or run harness.
| Model configuration | Score | Harness and setup | Source and date |
|---|---|---|---|
| GPT-6 Cyber (Effort unspecified)Approved Daybreak Red organizations | 48.45% | Harness and effort not specified | OpenAI, 2026-11-05; System card section 7.0.5; precision retained as published |
| GPT-6 Astra (Effort unspecified) | 42.43% | Harness and effort not specified | OpenAI, 2026-11-05; System card section 7.0.5; precision retained as published |
Hardened evaluation
The same card's section 7.0.6 reports a separate hardened variant. It adds stronger userspace exploit mitigations, enables the V8 sandbox, and uses stricter kernel configurations where supported. These scores measure a different setting from standard ExploitGym.
| Model configuration | Score | Harness and setup | Source and date |
|---|---|---|---|
| GPT-6 Cyber (Effort unspecified)Approved Daybreak Red organizations | 36.94% | Stronger userspace, V8 sandbox, and kernel mitigations | OpenAI, 2026-11-05; System card section 7.0.6; distinct task setting |
| GPT-6 Astra (Effort unspecified) | 29.69% | Stronger userspace, V8 sandbox, and kernel mitigations | OpenAI, 2026-11-05; System card section 7.0.6; distinct task setting |
Google Carbon release result
Google's Carbon evaluation methodology specifies Antigravity, one attempt per task, no internet access, and benchmark mitigations disabled. Its October 28 launch result reflects the Fairwind evaluation configuration with guardrails disabled. The methodology discusses both six-hour and ten-hour timeouts without identifying which produced the displayed table value.
| Model configuration | Score | Harness and setup | Source and date |
|---|---|---|---|
| Gemini 4 Carbon (High) | 54.8% | Guardrails disabled; Fairwind evaluation configuration | Google, 2026-10-28; Google methodology; source discusses 6-hour and 10-hour timeouts without tying the table value to one |
How to compare these results
Compare models within a source's published evaluation setup. Differences in task revisions, time limits, harnesses, mitigations, and access configurations can change what a score measures. ExploitGym Hardened and the scope-following ExploitGym Honeypot evaluation are separate settings. ExploitBench, SEC-bench Pro, and SRE-Bench test different tasks.