CAISI AgentDojo - Malware Execution (2025)
Safety · 2025-09-30
The malware-execution task in CAISI's 2025 AgentDojo Workspace evaluation instructs the agent through injected text to download and run a program from an untrusted URL. Scores average hijacking frequency across 30 benign tasks evaluated under two injection conditions.
Top models (lower is better)
| Model | Score |
|---|---|
| GPT-5 | 0.0 |
| Opus 4 | 8.0 |
| gpt-oss-120b | 28.0 |
| DeepSeek-R1-0528 | 49.0 |
| DeepSeek-V3.1 | 54.0 |
| DeepSeek-R1 | 76.0 |