Anthropic Malicious Computer Use - Refusal Rate: leaderboard
Anthropic's malicious computer-use evaluation: refusal rate on 112 harmful tasks (surveillance and data collection, harmful content, scaled abuse) given GUI and CLI tools in a sandbox (percent, higher is better). From Section 5.1.2 of the Claude Fable 5.1 & Claude Mythos 5.1 System Card (September 2026), task set unchanged since the Opus 5 card; older-model rows come from the Opus 5 and Sonnet 5 cards.
Metric: Refusal Rate (%). Source: www-cdn.anthropic.com. Saturation forecast: Around February 2027. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 5 | 93.75 | #16 |
| 2 | Claude Mythos Preview | 93.75 | #5 |
| 3 | Claude Mythos 5.1 | 87.5 | #2 |
| 4 | Claude Mythos 5 | 85.71 | #8 |
| 5 | Claude Sonnet 4.6 | 84.82 | #94 |
| 6 | Claude Sonnet 5 | 84.68 | #65 |
| 7 | Claude Opus 4.8 | 81.7 | #39 |
| 8 | Claude Opus 5.5 | 79.46 | #3 |
No result here: #6 GPT-6 Astra, #7 GPT-6.1 Sol, #9 Gemini 4 Argon.
Interactive version: theaggregate.ai/benchmark?slug=anthropic-malicious-computer-use-refusal-rate · How It Works · Data refreshed daily, snapshot 2026-11-05.