Tulu 3 Safety Eval
OLMo 2 1B Instruct 87.6% leads since Apr 2025 Fading Models answer harmful, jailbreak and benign prompts from several safety benchmarks in Ai2's safety-eval suite. A WildGuard classifier judges the outputs, and the score is an average.
An aggregate used in the Tulu 3 release. The member benchmarks include WildGuardTest, HarmBench and XSTest. Released month taken from the Tulu 3 release.
- Level
- Everyday Most adults, without training
- Scored
- Graded by a model A language model scores each response, usually against a rubric.
- Answers
- Open-ended Free-form text, documents, designs or media judged as a whole.
OLMo 2 1B Instruct leads with 87.6% among models released by Apr 2026
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.
The newest model with a score here came out in Apr 2026, and nobody has reported a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:
- GPT-6 Astra
- Claude Opus 5.5
- GPT-6.1 Terra
- GPT-6.1 Sol
- and 16 more
- 1 OLMo 2 1B Instruct 87.6%
- 2 Tulu 3 405B 86.7%
- 3 OLMo 2 32B Instruct 85.9%
- 4 Olmo 3.1 32B Think 83.6%
- 5 Granite 4.1 30B 78.2%
- 6 Olmo 3 32B Think 68.8%
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | OLMo 2 1B Instruct | 87.6% | Apr 2025 | Model card | |
| 2 | Tulu 3 405B | 86.7% | Jan 2025 | Model card | |
| 3 | OLMo 2 32B Instruct | 85.9% | Mar 2025 | Model card | |
| 4 | Olmo 3.1 32B Think | 83.6% | Dec 2025 | Model card | |
| 5 | Granite 4.1 30B | 78.2% | Apr 2026 | Model card | |
| 6 | Olmo 3 32B Think | 68.8% | Nov 2025 | Model card |
2 models have held number one since its release in Nov 2024
Lab claims count from the day the card came out, independent runs from the model’s release.
The best score passed 85% in Nov 2024
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Labs headline it less often now, last in Apr 2026
Launch posts from every lab that has listed it, one tick per post. Tall colored ticks include it, and short gray ones leave it out.
| Date | Launch post | Score | As the post writes it |
|---|---|---|---|
| 29 Apr 2026 | Granite 4.1 30B | 78.2 | Tulu3 Safety Eval Avg |
| 12 Dec 2025 | Olmo 3.1 32B Think | 83.6 | Safety (Ai2 safety suite average) |
| 20 Nov 2025 | Olmo 3 32B Think | 68.8 | Safety (Ai2 safety suite average) |
| 29 Apr 2025 | OLMo 2 1B Instruct | 87.6 | Safety (Ai2 safety suite average) |
| 13 Mar 2025 | OLMo 2 32B Instruct | 85.9 | Safety (Ai2 safety suite average) |
| 30 Jan 2025 | Tulu 3 405B | 86.7 | Safety (6 task avg.) |
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 |
|---|---|---|---|---|
| 15 Dec 2025 | Olmo 3 32B Think technical report | 68.8 | Safety, average of the Ai2 safety suite, Final Think 3.0 column (Table 14) | Safety |
| 15 Dec 2025 | Olmo 3.1 32B Think technical report | 83.6 | Safety, average of the Ai2 safety suite (Table 14) | Safety |
| 8 Oct 2025 | OLMo 2 1B Instruct technical report | 87.6 | Safety (Table 7, Table 23) | Safety |
| 31 Dec 2024 | OLMo 2 32B Instruct technical report | 85.9 | Safety, Tulu 3 Safety average (Tables 7 and 16) | Safety |
| 22 Nov 2024 | Tulu 3 405B technical report | 86.7 | Safety (6 task avg.), RLVR (Table 4) | 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%