Red Teaming Network pairwise safety comparison
Red teamers compare anonymized responses from two models in free-form risky conversations. Scored as the share of ratings where a model is safer.
Human red teamers rate only conversations with at least one perceived unsafe response.
- Level
- Expert Specialists with years of training, often over hours of work
- Scored
- Ranked by votes People pick the better of two answers, as in an arena, and the votes are turned into ratings.
- Answers
- Open-ended Free-form text, documents, designs or media judged as a whole.
o1 leads with 59.8%
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 o1 59.8%
- 2 OpenAI deep research 59.2%
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | o1 | 59.8% | Dec 2024 | Model card | |
| 2 | OpenAI deep research | 59.2% | Feb 2025 | Model card |
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.
3 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 |
|---|---|---|---|---|
| 25 Feb 2025 | OpenAI deep research system card | 59.2 | Targeted red teaming for risky advice, deep research ranked safer than GPT-4o 59.2% ± 3.9%, vs o3-mini 45.4% (Table 7) | Safety |
| 31 Jan 2025 | o3-mini system card | Pairwise Safety Comparison, win rates vs GPT-4o and o1 by self, peer and third-party raters, o3-mini-near-final-checkpoint, no overall score (Table 8) | Safety | |
| 5 Dec 2024 | o1 system card | 59.8 | Pairwise Safety Comparison, o1 rated safer vs gpt-4o by red teamers (Table 9) | Safety |
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%