DTap-Bench (Direct Attacks) - macOS: leaderboard
Direct misuse of tool-using agents in the macOS desktop environments of DTap-Bench (Chen et al., 2026): the user is the attacker and pursues a harmful goal with optimized prompt, tool, skill and environment injections. Attack success rate in percent over 50 tasks, averaged over risk categories and judged by a rule-based check of the final environment state; lower is better.
Metric: Attack success rate (%; 50 direct red-teaming macOS desktop tasks: the user is the attacker, with prompt, tool, skill and environment injections; an attack succeeds when a rule-based check of the final environment state finds the malicious goal fully achieved; fixed attack sequences that DTap-Red optimized on a surrogate agent; mean of its risk-category rates). Source: arxiv.org. Saturation forecast: Around February 2027. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 6.3 | #48 |
| 2 | DeepSeek V4 Pro | 10.4 | #181 |
| 3 | Claude Sonnet 4.5 | 18.3 | #244 |
| 4 | GPT-OSS-120B | 20 | #699 |
| 5 | GPT-5.2 | 26.2 | #172 |
| 6 | GPT-5.4 | 30.4 | #121 |
| 7 | Gemini 3 Pro | 30.4 | #135 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=dtap-bench-direct-attacks-macos · How It Works · Data refreshed daily, snapshot 2026-11-12.