ASPI
Agents complete simulated tasks with incomplete instructions. The authors compare adherence to user intent across matched versions of each task.
The paper reports results for ten models on 728 scenarios. Its published measure is attack success, so lower is better.
- Made by
- Scale AI
- Released
- May 2026
- Kind
- Deception and alignment
- Level
- Everyday
- Scored
- Checked automatically
- Answers
- Agent task
- Tasks
- All public
- Measures
- Propensity
- Risk
- Cyber offence
- Behavior
- Misaligned with intent
- Language
- English
- Interaction
- Agentic
- Environment
- APIs and tools, A simulator
- Items from
- Written by experts
- Refresh
- One fixed set
- Run by
- Anyone
- Strongest model tested
- GPT-5.5 #21 on the capability index
- Last model added
- April 2026
- Scores refreshed
- November 20th, 2026
Gemini 3.1 Pro leads with 17.8% among models released by April 2026
Best score per model from the benchmark's paper, read on November 20th, 2026. A paper’s table doesn’t add newer models. The headline averages attack success across six conditions, giving each condition equal weight. Where a model ran with several settings or harnesses, we keep its best.
The newest model with a score here came out in April 2026, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:
- GPT-6.2 Astra
- Claude Fable 5.2
-
Gemini 4 Carbon
- GPT-6 Astra
- and 16 more
- 1
Gemini 3.1 Pro Prompt guard 17.8%
- 2
Gemini 3 Flash Prompt guard 21.3%
- 3 o3 No defense 25.6%
- 4 GPT-5.5 No defense 26.9%
- 5 GPT-5.4 No defense 31.7%
- 6
Qwen3-32B No defense 33.7%
- 7 Claude Opus 4.7 No defense 34.3%
- 8 Kimi K2.5 No defense 51.9%
- 9
Qwen3-235B-A22B-Instruct Jul 2025, No defense 53.3%
- 10
DeepSeek-V3.2 No defense 68.3%
Source: The ASPI paper
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | | 17.8% | Prompt guard | Feb 2026 | Official leaderboard |
| 2 | | 21.3% | Prompt guard | Dec 2025 | Official leaderboard |
| 3 | o3 | 25.6% | No defense | Apr 2025 | Official leaderboard |
| 4 | GPT-5.5 | 26.9% | No defense | Apr 2026 | Official leaderboard |
| 5 | GPT-5.4 | 31.7% | No defense | Mar 2026 | Official leaderboard |
| 6 | | 33.7% | No defense | Apr 2025 | Official leaderboard |
| 7 | Claude Opus 4.7 | 34.3% | No defense | Apr 2026 | Official leaderboard |
| 8 | Kimi K2.5 | 51.9% | No defense | Jan 2026 | Official leaderboard |
| 9 | | 53.3% | No defense | Jul 2025 | Official leaderboard |
| 10 | | 68.3% | No defense | Dec 2025 | Official leaderboard |
Rankings by part
We read 30 parts from the benchmark’s paper, each ranking models on one part of the test, such as exec_tool, exec_user and exec_next. Claude Opus 4.7 leads 12 of them, and 5 other models lead the rest. Arrows show how far a model moved from its overall rank among the models we track.
exec_tool
10 models we track have a score here.
- 1
Gemini 3.1 Pro 0%
- 2 GPT-5.4 ↑3 0%
- 3 GPT-5.5 ↑1 0%
- 4
Gemini 3 Flash ↓2 0.30%
- 5 o3 ↓2 1.80%
- 6 Claude Opus 4.7 ↑1 2.10%
- 7 Kimi K2.5 ↑1 11.1%
- 8
Qwen3-32B ↓2 18.7%
- 9
Qwen3-235B-A22B-Instruct Jul 2025 41.9%
- 10
DeepSeek-V3.2 65.4%
exec_user
10 models we track have a score here.
- 1
Qwen3-32B ↑5 29.8%
- 2
Gemini 3 Flash 30.1%
- 3
Gemini 3.1 Pro ↓2 30.4%
- 4 o3 ↓1 33.5%
- 5
Qwen3-235B-A22B-Instruct Jul 2025 ↑4 47%
- 6 GPT-5.5 ↓2 54%
- 7
DeepSeek-V3.2 ↑3 65.5%
- 8 GPT-5.4 ↓3 66.3%
- 9 Kimi K2.5 ↓1 68.3%
- 10 Claude Opus 4.7 ↓3 75.7%
exec_next
10 models we track have a score here.
- 1
Gemini 3.1 Pro 29.3%
- 2
Gemini 3 Flash 29.8%
- 3 o3 35.2%
- 4
Qwen3-32B ↑2 41.3%
- 5 GPT-5.5 ↓1 50.4%
- 6 GPT-5.4 ↓1 52.5%
- 7
Qwen3-235B-A22B-Instruct Jul 2025 ↑2 58.4%
- 8 Claude Opus 4.7 ↓1 64.8%
- 9 Kimi K2.5 ↓1 66.3%
- 10
DeepSeek-V3.2 72.2%
clarif_tool
10 models we track have a score here.
- 1 GPT-5.4 ↑4 0%
- 2
Gemini 3.1 Pro ↓1 0.20%
- 3 GPT-5.5 ↑1 0.20%
- 4
Gemini 3 Flash ↓2 1.30%
- 5 o3 ↓2 1.50%
- 6 Claude Opus 4.7 ↑1 2.60%
- 7 Kimi K2.5 ↑1 12.2%
- 8
Qwen3-32B ↓2 14.5%
- 9
Qwen3-235B-A22B-Instruct Jul 2025 36.8%
- 10
DeepSeek-V3.2 50.9%
clarif_ask_user
10 models we track have a score here.
- 1 Claude Opus 4.7 ↑6 2.20%
- 2 GPT-5.5 ↑2 4.90%
- 3 GPT-5.4 ↑2 12%
- 4
Gemini 3.1 Pro ↓3 12.3%
- 5
Gemini 3 Flash ↓3 23.9%
- 6 o3 ↓3 34%
- 7
Qwen3-32B ↓1 42.2%
- 8 Kimi K2.5 63.1%
- 9
Qwen3-235B-A22B-Instruct Jul 2025 63.2%
- 10
DeepSeek-V3.2 77.3%
clarif_user
10 models we track have a score here.
- 1
Gemini 3.1 Pro 34.3%
- 2
Gemini 3 Flash 37.3%
- 3 o3 47.8%
- 4 GPT-5.5 51.6%
- 5
Qwen3-32B ↑1 55.7%
- 6 Claude Opus 4.7 ↑1 58.3%
- 7 GPT-5.4 ↓2 59.2%
- 8
Qwen3-235B-A22B-Instruct Jul 2025 ↑1 72.5%
- 9
DeepSeek-V3.2 ↑1 78.2%
- 10 Kimi K2.5 ↓2 90.6%
Utility: U_benign
10 models we track have a score here.
- 1 Claude Opus 4.7 ↑6 97.8%
- 2
Gemini 3.1 Pro ↓1 95.6%
- 3
Gemini 3 Flash ↓1 92.6%
- 4 GPT-5.4 ↑1 88.5%
- 5 GPT-5.5 ↓1 86.4%
- 6
DeepSeek-V3.2 ↑4 80.1%
- 7 Kimi K2.5 ↑1 68.8%
- 8
Qwen3-235B-A22B-Instruct Jul 2025 ↑1 64.1%
- 9
Qwen3-32B ↓3 45.6%
- 10 o3 ↓7 44%
Utility: U_tool
10 models we track have a score here.
- 1 Claude Opus 4.7 ↑6 93.7%
- 2
Gemini 3.1 Pro ↓1 91.2%
- 3
Gemini 3 Flash ↓1 90.1%
- 4 GPT-5.4 ↑1 84.6%
- 5 GPT-5.5 ↓1 81.5%
- 6 Kimi K2.5 ↑2 64.8%
- 7 o3 ↓4 42.4%
- 8
Qwen3-32B ↓2 29%
- 9
DeepSeek-V3.2 ↑1 28.2%
- 10
Qwen3-235B-A22B-Instruct Jul 2025 ↓1 26.9%
Utility: U_user
10 models we track have a score here.
- 1
Gemini 3.1 Pro 44.1%
- 2 Claude Opus 4.7 ↑5 38.9%
- 3 GPT-5.5 ↑1 38.6%
- 4
Gemini 3 Flash ↓2 37.4%
- 5 GPT-5.4 36%
- 6 Kimi K2.5 ↑2 28.7%
- 7
DeepSeek-V3.2 ↑3 27.8%
- 8
Qwen3-235B-A22B-Instruct Jul 2025 ↑1 23.5%
- 9 o3 ↓6 22.1%
- 10
Qwen3-32B ↓4 17.3%
Utility: U_next
10 models we track have a score here.
- 1 Claude Opus 4.7 ↑6 97.8%
- 2
Gemini 3.1 Pro ↓1 94.9%
- 3
Gemini 3 Flash ↓1 91.6%
- 4 GPT-5.5 87%
- 5 GPT-5.4 86.7%
- 6
DeepSeek-V3.2 ↑4 78.7%
- 7 Kimi K2.5 ↑1 70.3%
- 8
Qwen3-235B-A22B-Instruct Jul 2025 ↑1 64.1%
- 9
Qwen3-32B ↓3 43.7%
- 10 o3 ↓7 43.3%
Utility: U_clBen
10 models we track have a score here.
- 1 Claude Opus 4.7 ↑6 84%
- 2 GPT-5.5 ↑2 79.4%
- 3 Kimi K2.5 ↑5 78.9%
- 4
Gemini 3.1 Pro ↓3 78.7%
- 5 GPT-5.4 72%
- 6
DeepSeek-V3.2 ↑4 59.1%
- 7 o3 ↓4 56.9%
- 8
Gemini 3 Flash ↓6 55.2%
- 9
Qwen3-235B-A22B-Instruct Jul 2025 47.8%
- 10
Qwen3-32B ↓4 23.9%
Utility: U_clTool
10 models we track have a score here.
- 1 Claude Opus 4.7 ↑6 77.5%
- 2 GPT-5.5 ↑2 75.6%
- 3
Gemini 3.1 Pro ↓2 73.9%
- 4 Kimi K2.5 ↑4 71.5%
- 5 GPT-5.4 69.2%
- 6 o3 ↓3 53.2%
- 7
Gemini 3 Flash ↓5 52.5%
- 8
DeepSeek-V3.2 ↑2 34.5%
- 9
Qwen3-235B-A22B-Instruct Jul 2025 29.5%
- 10
Qwen3-32B ↓4 20.6%
Utility: U_clAsk
10 models we track have a score here.
- 1 Claude Opus 4.7 ↑6 82.9%
- 2 GPT-5.5 ↑2 77.3%
- 3
Gemini 3.1 Pro ↓2 69.3%
- 4 GPT-5.4 ↑1 67.8%
- 5 Kimi K2.5 ↑3 51.1%
- 6
Gemini 3 Flash ↓4 46%
- 7 o3 ↓4 31.8%
- 8
Qwen3-235B-A22B-Instruct Jul 2025 ↑1 24%
- 9
DeepSeek-V3.2 ↑1 23.1%
- 10
Qwen3-32B ↓4 14%
Utility: U_clUser
10 models we track have a score here.
- 1 Claude Opus 4.7 ↑6 75%
- 2 GPT-5.5 ↑2 63.1%
- 3
Gemini 3.1 Pro ↓2 57.8%
- 4 Kimi K2.5 ↑4 55.2%
- 5 GPT-5.4 54.9%
- 6
Gemini 3 Flash ↓4 42.4%
- 7 o3 ↓4 37.9%
- 8
DeepSeek-V3.2 ↑2 35.2%
- 9
Qwen3-235B-A22B-Instruct Jul 2025 33.2%
- 10
Qwen3-32B ↓4 17.8%
Tool Execution workspace
10 models we track have a score here.
- 1 GPT-5.4 ↑4 0%
- 2 Claude Opus 4.7 ↑5 0%
- 3 GPT-5.5 ↑1 0%
- 4
Gemini 3 Flash ↓2 0.20%
- 5 o3 ↓2 0.50%
- 6
Gemini 3.1 Pro ↓5 0.50%
- 7 Kimi K2.5 ↑1 2.90%
- 8
Qwen3-32B ↓2 10.3%
- 9
Qwen3-235B-A22B-Instruct Jul 2025 34.9%
- 10
DeepSeek-V3.2 55.5%
Tool Execution slack
10 models we track have a score here.
- 1
Gemini 3.1 Pro 0%
- 2 GPT-5.4 ↑3 0%
- 3 Claude Opus 4.7 ↑4 0%
- 4 GPT-5.5 0%
- 5
Gemini 3 Flash ↓3 7.20%
- 6 o3 ↓3 8.20%
- 7 Kimi K2.5 ↑1 41.2%
- 8
Qwen3-32B ↓2 57.7%
- 9
Qwen3-235B-A22B-Instruct Jul 2025 81.4%
- 10
DeepSeek-V3.2 90.7%
Tool Execution travel
10 models we track have a score here.
- 1 GPT-5.4 ↑4 0%
- 2 GPT-5.5 ↑2 0%
- 3 o3 1.10%
- 4
Gemini 3.1 Pro ↓3 1.10%
- 5
Gemini 3 Flash ↓3 3.20%
- 6 Kimi K2.5 ↑2 6.40%
- 7 Claude Opus 4.7 16%
- 8
Qwen3-32B ↓2 20.2%
- 9
Qwen3-235B-A22B-Instruct Jul 2025 31.9%
- 10
DeepSeek-V3.2 74.5%
Tool Execution banking
10 models we track have a score here.
- 1 GPT-5.4 ↑4 0%
- 2 Claude Opus 4.7 ↑5 0%
- 3 GPT-5.5 ↑1 0%
- 4 o3 ↓1 1.50%
- 5
Gemini 3 Flash ↓3 3.90%
- 6
Gemini 3.1 Pro ↓5 3.90%
- 7
Qwen3-32B ↓1 14.6%
- 8 Kimi K2.5 17.7%
- 9
Qwen3-235B-A22B-Instruct Jul 2025 41.5%
- 10
DeepSeek-V3.2 70.8%
Tool Clarification workspace
10 models we track have a score here.
- 1 Claude Opus 4.7 ↑6 0%
- 2 GPT-5.5 ↑2 5.10%
- 3 GPT-5.4 ↑2 11.2%
- 4
Gemini 3.1 Pro ↓3 14.1%
- 5
Qwen3-32B ↑1 29%
- 6
Gemini 3 Flash ↓4 29.4%
- 7 o3 ↓4 33.8%
- 8
Qwen3-235B-A22B-Instruct Jul 2025 ↑1 49.1%
- 9 Kimi K2.5 ↓1 49.2%
- 10
DeepSeek-V3.2 67.5%
Tool Clarification slack
10 models we track have a score here.
- 1 Claude Opus 4.7 ↑6 0%
- 2 GPT-5.5 ↑2 10.4%
- 3 GPT-5.4 ↑2 18.2%
- 4 o3 ↓1 37.5%
- 5
Gemini 3.1 Pro ↓4 43.3%
- 6
Gemini 3 Flash ↓4 65.7%
- 7
Qwen3-32B ↓1 78.2%
- 8
Qwen3-235B-A22B-Instruct Jul 2025 ↑1 85.7%
- 9 Kimi K2.5 ↓1 86.1%
- 10
DeepSeek-V3.2 97.6%
Tool Clarification travel
10 models we track have a score here.
- 1 GPT-5.5 ↑3 5.30%
- 2 Claude Opus 4.7 ↑5 10.9%
- 3 GPT-5.4 ↑2 12.8%
- 4
Gemini 3.1 Pro ↓3 37.6%
- 5
Gemini 3 Flash ↓3 51.1%
- 6 Kimi K2.5 ↑2 58.6%
- 7
Qwen3-32B ↓1 60.7%
- 8 o3 ↓5 61%
- 9
Qwen3-235B-A22B-Instruct Jul 2025 62.5%
- 10
DeepSeek-V3.2 80.5%
Tool Clarification banking
10 models we track have a score here.
- 1 Claude Opus 4.7 ↑6 0%
- 2 GPT-5.5 ↑2 0.80%
- 3 GPT-5.4 ↑2 9.10%
- 4 o3 ↓1 15.9%
- 5
Gemini 3 Flash ↓3 21.9%
- 6
Gemini 3.1 Pro ↓5 22.3%
- 7
Qwen3-32B ↓1 58.2%
- 8 Kimi K2.5 65.7%
- 9
DeepSeek-V3.2 ↑1 84.4%
- 10
Qwen3-235B-A22B-Instruct Jul 2025 ↓1 88.2%
User Execution workspace
10 models we track have a score here.
- 1
Qwen3-32B ↑5 25.3%
- 2 o3 ↑1 25.6%
- 3
Qwen3-235B-A22B-Instruct Jul 2025 ↑6 40.3%
- 4 GPT-5.5 44.7%
- 5 Kimi K2.5 ↑3 48.9%
- 6
DeepSeek-V3.2 ↑4 58.7%
- 7 GPT-5.4 ↓2 61.2%
- 8
Gemini 3.1 Pro ↓7 61.9%
- 9 Claude Opus 4.7 ↓2 65.4%
- 10
Gemini 3 Flash ↓8 76.2%
User Execution slack
10 models we track have a score here.
- 1 o3 ↑2 38.1%
- 2
Qwen3-32B ↑4 69.1%
- 3
Qwen3-235B-A22B-Instruct Jul 2025 ↑6 75.3%
- 4 GPT-5.5 88.7%
- 5
DeepSeek-V3.2 ↑5 94.8%
- 6 GPT-5.4 ↓1 94.8%
- 7
Gemini 3.1 Pro ↓6 96.9%
- 8 Claude Opus 4.7 ↓1 96.9%
- 9
Gemini 3 Flash ↓7 100%
- 10 Kimi K2.5 ↓2 100%
User Execution travel
10 models we track have a score here.
- 1
Qwen3-32B ↑5 23.4%
- 2 o3 ↑1 41.5%
- 3
Qwen3-235B-A22B-Instruct Jul 2025 ↑6 43.6%
- 4 GPT-5.5 69.2%
- 5 GPT-5.4 75.5%
- 6 Claude Opus 4.7 ↑1 84%
- 7
DeepSeek-V3.2 ↑3 85.1%
- 8
Gemini 3.1 Pro ↓7 86.2%
- 9 Kimi K2.5 ↓1 93.6%
- 10
Gemini 3 Flash ↓8 96.8%
User Execution banking
10 models we track have a score here.
- 1
Qwen3-32B ↑5 19.2%
- 2 GPT-5.5 ↑2 46.2%
- 3 o3 49.2%
- 4
Qwen3-235B-A22B-Instruct Jul 2025 ↑5 49.2%
- 5
DeepSeek-V3.2 ↑5 50.8%
- 6 GPT-5.4 ↓1 54.6%
- 7 Claude Opus 4.7 86.2%
- 8 Kimi K2.5 86.9%
- 9
Gemini 3 Flash ↓7 90%
- 10
Gemini 3.1 Pro ↓9 90%
User Clarification workspace
10 models we track have a score here.
- 1 Claude Opus 4.7 ↑6 38.6%
- 2 GPT-5.5 ↑2 44.1%
- 3
Qwen3-32B ↑3 45.5%
- 4 o3 ↓1 48.9%
- 5 GPT-5.4 54.4%
- 6
Qwen3-235B-A22B-Instruct Jul 2025 ↑3 60.5%
- 7
DeepSeek-V3.2 ↑3 66.4%
- 8
Gemini 3.1 Pro ↓7 67.7%
- 9
Gemini 3 Flash ↓7 73.2%
- 10 Kimi K2.5 ↓2 81.2%
User Clarification slack
10 models we track have a score here.
- 1 o3 ↑2 54.7%
- 2 GPT-5.5 ↑2 75.3%
- 3
Qwen3-32B ↑3 80%
- 4 Claude Opus 4.7 ↑3 81.9%
- 5
Qwen3-235B-A22B-Instruct Jul 2025 ↑4 83.9%
- 6 GPT-5.4 ↓1 85.2%
- 7
Gemini 3.1 Pro ↓6 96.7%
- 8
Gemini 3 Flash ↓6 98.5%
- 9 Kimi K2.5 ↓1 98.7%
- 10
DeepSeek-V3.2 100%
User Clarification travel
10 models we track have a score here.
- 1 o3 ↑2 68.8%
- 2 GPT-5.5 ↑2 69.2%
- 3 Claude Opus 4.7 ↑4 77.2%
- 4
Qwen3-32B ↑2 78.6%
- 5 GPT-5.4 83%
- 6
Qwen3-235B-A22B-Instruct Jul 2025 ↑3 85.4%
- 7
DeepSeek-V3.2 ↑3 87%
- 8 Kimi K2.5 92%
- 9
Gemini 3.1 Pro ↓8 95.7%
- 10
Gemini 3 Flash ↓8 96.6%
User Clarification banking
10 models we track have a score here.
- 1 o3 ↑2 30.2%
- 2 GPT-5.4 ↑3 33.1%
- 3 GPT-5.5 ↑1 37.7%
- 4 Claude Opus 4.7 ↑3 57.7%
- 5
Qwen3-32B ↑1 67.3%
- 6
DeepSeek-V3.2 ↑4 85.4%
- 7
Qwen3-235B-A22B-Instruct Jul 2025 ↑2 90.8%
- 8 Kimi K2.5 94.3%
- 9
Gemini 3.1 Pro ↓8 97.3%
- 10
Gemini 3 Flash ↓8 97.5%
Number one over time
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: The ASPI paper
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Official ASPI leaderboard
No lab model card reports it yet
We read the headline results of 540 launch posts from 57 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%
- Covert misbehavior (Thinking Machines) In use Agentic tasks that check whether a model misbehaves and conceals it. Scored by the rate of covert misbehavior.
Inkling 2 Mini 8.30%
- Destructive actions (Thinking Machines) In use Agentic tasks that check whether a model takes destructive actions it was not authorized to take. Scored by the rate of destructive actions.
Inkling 2 Small 4.70%
- Misbehavior (Thinking Machines) In use Agentic tasks that check whether a model fabricates results, ignores instructions or hides what it did when a task is hard. Scored by the rate of misbehavior.
Inkling 2 Mini 9.80%
Data audit
8 of 9 checks pass, 1 warning. Look first at the Newest score check. The newest came 7 months ago, in April 2026, for GPT-5.5. Every benchmark
Catalog record 2 pass
- Pass Task access All tasks are public.
- Pass Confidence Checked against a primary source.
Scores 1 warning 2 pass 2 notes
- Pass Independent runs Scores from the benchmark’s paper (10 entries).
- Pass Models scored 10 models we track.
- Warning Newest score The newest came 7 months ago, in April 2026, for GPT-5.5.
- Note Frontier models None of the top 10 has a score. The strongest tested is GPT-5.5 (#21).
- Note Headroom Lower is better, so there’s no ceiling to reach.
Leaderboard import 3 pass
- Pass Importer We read the board on arxiv.org: 14 rows for models we track at the last read.
- Pass Last read November 20th, 2026.
- Pass Read worked The last read worked.
Consistency 1 pass 1 note
- Note Lab claims No model has both a lab number and an independent run.
- Pass Known issues No published audit found problems.