PASB - Combined Attack: leaderboard
Metric: Attack success rate (%) of the combined attack (escape characters, a fake completion and a context-ignoring instruction combined): share of trials in which the agent calls the attacker's target skill (PASB indirect prompt injection on the deployed OpenClaw personal agent: for each of 131 risky skills from the OpenClaw skill registry a fixed benign query is answered, the step-1 tool observation carries the injected instruction, and the attack succeeds when the agent calls the targeted skill in step 2 (a sandboxed, auditable testbed); no defense; mean over repeated trials); lower is better. Source: arxiv.org. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 2.5 7B Instruct | 52.7 | #846 |
| 2 | GPT-4o Mini | 61.9 | #588 |
| 3 | Llama 3.1 70B Instruct | 66.8 | #548 |
Interactive version: theaggregate.ai/benchmark?slug=pasb-combined-attack · How It Works · Data refreshed daily, snapshot 2026-10-11.