Sample-level Attack Success Rate on Code-context prompt injection
27.11ASRClaude-3.5-Haiku
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Claude-3.5-HaikuVictim Model=Claude-3.5-Haiku, Defense Status=Before2026.06 | 27.11 | — | |
| Gemini-3.1-Flash-liteVictim Model=Gemini-3.1-Flash-lite, Defense Status=Before2026.06 | 24.14 | — | |
| GPT-5.1-Codex-miniVictim Model=GPT-5.1-Codex-mini, Defense Status=Before2026.06 | 18.32 | — | |
| CodeSentinelVictim Model=Gemini-3.1-Flash-lite, Defense Status=After2026.06 | 8.28 | -15.86 | |
| CodeSentinelVictim Model=Claude-3.5-Haiku, Defense Status=After2026.06 | 7.72 | -19.39 | |
| CodeSentinelVictim Model=GPT-5.1-Codex-mini, Defense Status=After2026.06 | 5.81 | -12.51 |