Oracle Poisoning Susceptibility on Oracle Poisoning Scenario 3 L2 Real SDK tool-use prefix-free (N=30)
100TrustClaude Sonnet 4.6
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Claude Sonnet 4.6Provider=Anthropic, Classification=Inclusive2026.05 | 100 | 88.4 | |
| Claude Opus 4.5Provider=Anthropic2026.05 | 100 | 88.4 | |
| GPT-4oProvider=OpenAI2026.05 | 100 | 88.4 | |
| GPT-4.1Provider=OpenAI2026.05 | 100 | 88.4 | |
| GPT-5-miniProvider=OpenAI2026.05 | 100 | 88.4 | |
| GPT-5.1Provider=OpenAI2026.05 | 100 | 88.4 | |
| GPT-5.2Provider=OpenAI2026.05 | 100 | 88.4 | |
| Gemini 2.5 ProProvider=Google2026.05 | 100 | 88.4 | |
| Gemini 2.0 FlashProvider=Google2026.05 | 100 | 88.1 | |
| Claude Sonnet 4.6Provider=Anthropic, Classification=Strict2026.05 | 80 | 61.4 |