Machiavelli Deception (CAIS)
Safety · 2023-04-06
Machiavelli measures harmful and deceptive behavior by LLM agents in goal-oriented text-based games. SAFE reports its deception measurements as a CAIS AI Dashboard risk component, where lower scores indicate less deceptive behavior; 100 is the random-policy baseline and values above 100 are possible for more deceptive behavior.
Top models (lower is better)
| Model | Score |
|---|---|
| GPT-5 Nano | 79.4 |
| Claude Fable 5 | 80.2 |
| GPT-5.6 Terra | 80.4 |
| Opus 4.5 | 80.8 |
| Grok 4.1 Fast | 81.3 |
| Haiku 4.5 | 81.5 |
| Sonnet 4.5 | 81.6 |
| GPT-5.5 | 81.7 |
| GPT-5.6 Sol | 82.7 |
| DeepSeek-V4-Pro | 83.0 |
| Opus 4.8 | 83.7 |
| Grok 4.20 | 83.8 |
| GPT-5 | 83.8 |
| Kimi K2 Thinking | 84.2 |
| GPT-5 Mini | 84.2 |
| Opus 4.6 | 84.4 |
| Grok 4.3 | 84.7 |
| Sonnet 4.6 | 84.9 |
| DeepSeek-R1 | 85.1 |
| Opus 4.7 | 85.7 |
| GPT-5.2 | 86.5 |
| GLM-5.1 | 86.7 |
| GPT-5.6 Luna | 87.9 |
| Muse Spark 1.1 | 88.2 |
| GPT-5.4 Nano | 88.5 |
| Gemini 3.1 Pro Preview | 89.2 |
| Gemini 2.5 Flash | 89.4 |
| DeepSeek-V3.2 | 89.6 |
| GPT-5.1 | 89.6 |
| Grok 4.5 | 90.9 |
| Kimi K2.6 | 92.1 |
| Gemini 3.5 Flash | 92.3 |
| GLM-5.2 | 92.3 |
| Kimi K3 | 92.6 |
| GPT-5.4 | 92.9 |
| Grok 4 | 93.4 |
| Kimi K2.5 | 93.5 |
| Gemini 2.5 Pro | 95.3 |
| Gemini 3 Flash Preview | 95.3 |
| GPT-5.4 Mini | 95.6 |