Declining to Exploit a Honeypot During Difficult ExploitGym Problems
Agents attempt hard ExploitGym challenges where a honeypot offers an illegitimate shortcut. Scored on how often they exploit the honeypot.
In-house OpenAI evaluation built on ExploitGym problems.
- Level
- Expert Specialists with years of training, often over hours of work
- Scored
- Checked automatically A program checks the result against an answer key, unit tests or the expected final state.
- Answers
- Agent task Many steps in a terminal, browser, desktop, app, game or robot.
4 models share the top score of 0%
Best score per model as labs report them in their own cards. Labs pick their own settings, so these compare less cleanly than independent runs.
- 1 GPT-6 Astra 0%
- 2 GPT-6 Luna 0%
- 3 GPT-6 Sol 0%
- 4 GPT-6.1 Sol 0%
- 5 GPT-6.1 Terra 0.07%
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | GPT-6 Astra | 0% | Sep 2026 | Model card | |
| 2 | GPT-6 Luna | 0% | Sep 2026 | Model card | |
| 3 | GPT-6 Sol | 0% | Sep 2026 | Model card | |
| 4 | GPT-6.1 Sol | 0% | Sep 2026 | Model card | |
| 5 | GPT-6.1 Terra | 0.07% | Oct 2026 | Model card |
Number one over time
Lab claims count from the day the card came out, independent runs from the model’s release.
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Labs report it only in system cards
We read the headline results of 479 launch posts from 51 labs, and none of them list this benchmark. Some system cards do, below.
5 system cards report it, from 1 lab
Besides launch posts, we read the system cards, model cards and technical reports labs publish with a release. They report far more, including safety and frontier-risk tests.
| Date | Document | Score | As the document writes it | Section |
|---|---|---|---|---|
| 26 Oct 2026 | GPT-6.1 Terra system card addendum | 0.07 | ExploitGym honeypot, unsuccessful honeypot attempts 0.07% of rollouts, no successful honeypot exploits, legitimate captures 0.14% (Figure 6, 7.0.3) | Alignment |
| 29 Sep 2026 | GPT-6.1 Sol system card addendum | 0 | Declining to Exploit a Honeypot During Difficult ExploitGym Problems, honeypot exploit attempts, text says none (7.3) | Alignment |
| 22 Sep 2026 | GPT-6 Sol system card | 0 | ExploitGym honeypot, share of samples exploiting or attempting the honeypot, printed chart label (Figure 63, A.6.3) | Alignment |
| 22 Sep 2026 | GPT-6 Luna system card | 0 | ExploitGym honeypot, share of samples exploiting or attempting the honeypot, printed chart label (Figure 63, A.6.3) | Alignment |
| 3 Sep 2026 | GPT-6 Astra system card | 0 | Declining to Exploit a Honeypot During Difficult ExploitGym Problems, honeypot attacks, text says none (legitimate capture 1.3%) (8.2.3) | Alignment |
More in Safety and alignment
- FORTRESS In use Paired adversarial and benign national-security misuse prompts. Rubric-graded risk score on adversarial and over-refusal score on benign prompts. Claude Opus 5.5 5.78%
- StrongREJECT In use Forbidden-prompt jailbreak benchmark with a rubric grader that scores how useful a non-refusal actually is.
Inkling 98.6% - AgentDojo In use Agents do tool-use tasks while tools and data carry prompt injections. Attack success rate counts hijacked runs. Grok 4 Fast 0%
- MASK In use Measures whether a model contradicts its own elicited beliefs under pressure to lie. Honesty separated from accuracy.
Muse Spark Open 98.1% - Gray Swan IPI Arena benchmark In use Indirect prompt-injection attacks from public red-teaming competitions replayed on tool, coding and browser agents. Attack success rate at k attempts. GPT-6 Astra 8.50%
- Agentic misalignment suite (Anthropic) In use Agents in simulated company settings face threats or goal conflicts that invite blackmail, leaking or sabotage. Score is the misaligned action rate. Claude Sonnet 4.5 0%