ASPI - execution-state tool-channel attack success rate
Safety · 2026-05-17
Scale AI's ASPI (Ambiguous-State Prompt Injection) builds 728 task-attack groups on AgentDojo's workspace, Slack, travel and banking suites, each pairing a benign user task with an attacker goal. In this condition (exec_tool), the standard AgentDojo setting, the agent gets the fully specified prompt and the attacker's instruction arrives in a task tool's return value; the attack success rate is the share of the 728 groups in which AgentDojo's deterministic security check finds the attacker goal achieved. Lower is better; it measures robustness to prompt injection, not capability.
Top models (lower is better)
| Model | Score |
|---|---|
| GPT-5.5 | 0.0 |
| GPT-5.4 | 0.0 |
| Gemini 3.1 Pro Preview | 1.1 |
| o3 | 1.8 |
| Opus 4.7 | 2.1 |
| Gemini 3 Flash Preview | 2.2 |
| Kimi K2.5 | 11.1 |
| Qwen3 32B | 18.7 |
| Qwen3 235B A22B Instruct 2507 | 41.9 |
| DeepSeek-V3.2 | 65.4 |