DTap-Bench (Indirect Attacks) - macOS: leaderboard
Indirect prompt injection against tool-using agents in the macOS desktop environments of DTap-Bench (Chen et al., 2026): third-party tool, skill or environment injections try to divert a benign user task to a malicious goal. Attack success rate in percent over 53 tasks, averaged over risk categories and judged by a rule-based check of the final environment state; lower is better.
Metric: Attack success rate (%; 53 indirect red-teaming macOS desktop tasks: third-party tool, skill and environment injections into a benign user task; 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: Estimated already saturated. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 0 | #48 |
| 2 | DeepSeek V4 Pro | 0 | #181 |
| 3 | GPT-OSS-120B | 18.3 | #699 |
| 4 | Claude Sonnet 4.5 | 20.4 | #244 |
| 5 | GPT-5.4 | 26.7 | #121 |
| 6 | GPT-5.2 | 26.7 | #172 |
| 7 | Gemini 3 Pro | 34.6 | #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-indirect-attacks-macos · How It Works · Data refreshed daily, snapshot 2026-11-12.