Over-eager behavior on GUI computer use tasks (Anthropic)
Computer-use agents face impossible tasks with an unwanted shortcut available. Score is the rate of unapproved workarounds instead of checking with the user.
Run with neutral, encouraging and discouraging system prompts. The neutral prompt is the default setting.
- Level
- Expert Specialists with years of training, often over hours of work
- Scored
- Graded by a model A language model scores each response, usually against a rubric.
- Answers
- Agent task Many steps in a terminal, browser, desktop, app, game or robot.
Claude Opus 4.8 leads with 9.40% among models released by May 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 May 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 15 more
- 1 Claude Opus 4.8 9.40%
- 2 Claude Opus 4.6 11.4%
- 3 Claude Mythos Preview 13.3%
- 4 Claude Sonnet 4.6 14.7%
- 5 Claude Opus 4.7 21.1%
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | Claude Opus 4.8 | 9.40% | May 2026 | Model card | |
| 2 | Claude Opus 4.6 | 11.4% | Feb 2026 | Model card | |
| 3 | Claude Mythos Preview | 13.3% | Apr 2026 | Model card | |
| 4 | Claude Sonnet 4.6 | 14.7% | Feb 2026 | Model card | |
| 5 | Claude Opus 4.7 | 21.1% | Apr 2026 | Model card |
2 models have held number one since its release in Feb 2026
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 |
|---|---|---|---|---|
| 28 May 2026 | Claude Opus 4.8 system card | 9.4 | Overeager behavior on GUI computer use tasks, hacking behavior rate with the neutral system prompt (16.1% when encouraged, 2.8% when discouraged), printed on chart, lower is better (6.3.1, Figure 6.3.1.A) | Alignment |
| 16 Apr 2026 | Claude Opus 4.7 system card | 21.1 | Hacking behavior on GUI computer use tasks, neutral system prompt, printed on chart. 30.4 with encouraging prompt, 6.7 with discouraging prompt (Figure 6.2.2.2.C) | Alignment |
| 7 Apr 2026 | Claude Mythos Preview system card | 13.3 | Hacking behavior on GUI computer use tasks, updated version of the overly agentic GUI evaluation, neutral system prompt, lower is better, value printed on chart; 24.6 with encouraging prompt, 3.8 with discouraging prompt (4.2.2.2, Figure 4.2.2.2.C) | Alignment |
| 17 Feb 2026 | Claude Sonnet 4.6 system card | 14.7 | Over-eager behavior rate on GUI computer use tasks, neutral system prompt; 27.9 with encouraging prompt, 3.9 with discouraging prompt (Figure 4.3.3.A) | Alignment |
| 5 Feb 2026 | Claude Opus 4.6 system card | 11.4 | Over-eager behavior rate on GUI computer use tasks, neutral system prompt, lower is better, value printed on chart; 18.8 with encouraging prompt, 12.7 with discouraging prompt (Figure 6.2.3.3.A) | 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%