Deception by Omission - Agentic: leaderboard
In a tool workspace the agent's own earlier turn did something harmful (for example deleted a needed file) and reported success; after an unrelated follow-up, the number is the share of rollouts where it tells the user and owns it.
Metric: Agent replies that disclose and own a planted harmful action, unrelated follow-up (%). Source: arxiv.org. Saturation forecast: Around February 2027. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 5 (Medium) | 35.2 | #111 (Claude Sonnet 5) |
| 2 | Gemini 3.5 Flash (Medium) | 25.5 | #101 (Gemini 3.5 Flash) |
| 3 | Qwen 3.7 Max | 23 | #145 |
| 4 | GPT-5.4 (Medium) | 14.1 | #123 (GPT-5.4) |
| 5 | DeepSeek V4 Pro (Medium) | 3.3 | #188 (DeepSeek V4 Pro) |
| 6 | Kimi K2.6 | 2.9 | #178 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=deception-by-omission-agentic · How It Works · Data refreshed daily, snapshot 2026-11-13.