Indirect Prompt Injection on LLM Behavior Subset 1
99.8IRActor-Critic
Evaluation Results
| Method | Links | |
|---|---|---|
| Actor-CriticMemory regime=Tool, Model=GPT-5.52026.05 | 99.8 | |
| Actor-CriticMemory regime=Tool, Model=GPT-5.42026.05 | 99.4 | |
| Actor-CriticMemory regime=Tool, Model=DeepSeek-v42026.05 | 96.2 | |
| Actor-CriticMemory regime=Tool, Model=Kimi-K2.62026.05 | 95 | |
| Actor-CriticMemory regime=Tool, Model=Gemini-3.12026.05 | 88.6 | |
| Actor-CriticMemory regime=External, Model=DeepSeek-v42026.05 | 86.4 | |
| Actor-CriticMemory regime=External, Model=GPT-5.42026.05 | 80.4 | |
| User ReviewMemory regime=Tool, Model=DeepSeek-v42026.05 | 79 | |
| Actor-CriticMemory regime=External, Model=Kimi-K2.62026.05 | 78.8 | |
| Actor-CriticMemory regime=External, Model=GPT-5.52026.05 | 75 | |
| Actor-CriticMemory regime=Tool, Model=Sonnet-4.62026.05 | 64.2 | |
| User ReviewMemory regime=Tool, Model=Kimi-K2.62026.05 | 62.8 | |
| User ReviewMemory regime=Tool, Model=Gemini-3.12026.05 | 62.4 | |
| Actor-CriticMemory regime=External, Model=Gemini-3.12026.05 | 54.2 | |
| Actor-CriticMemory regime=External, Model=Sonnet-4.62026.05 | 13.6 | |
| User ReviewMemory regime=External, Model=DeepSeek-v42026.05 | 5.4 | |
| User ReviewMemory regime=Tool, Model=GPT-5.52026.05 | 4.2 | |
| User ReviewMemory regime=External, Model=Gemini-3.12026.05 | 3.8 | |
| User ReviewMemory regime=Tool, Model=GPT-5.42026.05 | 3 | |
| User ReviewMemory regime=External, Model=GPT-5.52026.05 | 2.6 | |
| User ReviewMemory regime=External, Model=GPT-5.42026.05 | 2.2 | |
| User ReviewMemory regime=External, Model=Kimi-K2.62026.05 | 2 | |
| User ReviewMemory regime=External, Model=Sonnet-4.62026.05 | 0.2 | |
| User ReviewMemory regime=Tool, Model=Sonnet-4.62026.05 | 0 |