IPI Bench
Agents complete workplace workflows while attackers add misleading records to an emulated company with 11 services. The benchmark pairs six workflow types with harmful injected objectives across 34 tasks. The headline here measures single-query attack success as a percentage, where lower is better.
Models run at medium reasoning effort without added defenses. Judges assess task-specific rubrics, and reviewers check every claimed attack success. The report separately gives independent and iterative attack budgets. Clean-workflow cost and total tokens describe runs without attacks. The report dates itself to October 2026.
- Made by
- General Analysis
- Released
- October 2026
- Kind
- Jailbreaks and prompt injection
- Level
- Expert
- Scored
- Graded by a model
- Answers
- Agent task
- Tasks
- Kept private
- Strongest model tested
- GPT-6 Astra #3 on the capability index
- Last model added
- September 2026
- Scores refreshed
- November 12th, 2026
Claude Opus 5.5 leads with 4.80%
Best score per entry from the benchmark's own leaderboard, read on November 12th, 2026. Where a model ran with several settings or harnesses, we keep its best. Names in gray aren't in our model list: fine-tunes, agent systems, products and older models. They don't count toward leaders.
- 1 Claude Opus 5.5 medium effort; one-query attack 4.80%
- 2 GPT-6 Astra medium effort; one-query attack 7.50%
- 3
Gemini 4 Argon medium effort; one-query attack 12.6% - 4 GPT-6.1 Sol medium effort; one-query attack 18.4%
- 5 GPT-6 Luna medium effort; one-query attack 31.8%
- 6
DeepSeek V4.1 Flash medium effort; one-query attack 35.3% - 7 Muse Wonder 1.0 medium effort; one-query attack 38.5%
Source: Official IPI Bench leaderboard
| # | Entry | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | Claude Opus 5.5 | 4.80% | medium effort; one-query attack | Sep 2026 | Official leaderboard |
| 2 | GPT-6 Astra | 7.50% | medium effort; one-query attack | Sep 2026 | Official leaderboard |
| 3 | Gemini 4 Argon | 12.6% | medium effort; one-query attack | Sep 2026 | Official leaderboard |
| 4 | GPT-6.1 Sol | 18.4% | medium effort; one-query attack | Sep 2026 | Official leaderboard |
| 5 | GPT-6 Luna | 31.8% | medium effort; one-query attack | Sep 2026 | Official leaderboard |
| 6 | DeepSeek V4.1 Flash | 35.3% | medium effort; one-query attack | Sep 2026 | Official leaderboard |
| 7 | Muse Wonder 1.0 | 38.5% | medium effort; one-query attack | Official leaderboard |
Main objective score against single-query attack success
Agents receive the main objective score in separate runs without injected material. The line joins the runs that no other run beats on both measures. We redrew this chart from the numbers on the official leaderboard.
| Model | Setting | Main objective score without injected material | Attack success rate |
|---|---|---|---|
| GPT-6 Astra | medium effort; one-query attack | 98.1% | 7.5% |
| Claude Opus 5.5 | medium effort; one-query attack | 97% | 4.8% |
| GPT-6.1 Sol | medium effort; one-query attack | 95.6% | 18.4% |
| GPT-6 Luna | medium effort; one-query attack | 95.3% | 31.8% |
| Gemini 4 Argon | medium effort; one-query attack | 92.6% | 12.6% |
| Muse Wonder 1.0 | medium effort; one-query attack | 92.2% | 38.5% |
| DeepSeek V4.1 Flash | medium effort; one-query attack | 87.5% | 35.3% |
Number one over time
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: Official IPI Bench leaderboard
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Official IPI Bench leaderboard
No lab model card reports it yet
We read the headline results of 522 launch posts from 52 labs, and none of them list this benchmark.
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% - 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. Gemini 4 Carbon 0%
- DistressBench In use Models reply as a chatbot to 718 simulated users in psychological crisis, in single-turn and multi-turn chats. Three judge models score replies against clinician-written rubrics.
Muse Spark 1.2 88.4% - Lakera B3 (Backbone Breaker Benchmark) In use Crowdsourced prompt-injection attacks hit backbone LLMs inside agent threat snapshots at three defense levels. The score is attack resistance, 100 minus the board's vulnerability percentage. Claude Opus 4.5 89%
- IHEval In use Models face conflicts between system, user and tool instructions. Reports a composite score for following the higher-priority instruction. GPT-4o 70%