MemDrift - Risk Permissiveness: leaderboard

Metric: Deflection score under biased memory (1-5; Claude Opus 4.6 judge compares five tool calls against the no-memory calls, 1 = unchanged), direct memory injection, Risk Permissiveness bias (21 scenarios). Source: arxiv.org. Saturation forecast: Around July 2027. 7 models tracked.

Top models

#ModelScore
1GPT-5.2 (Medium)2.67
2GPT-5.4 (Medium)3.57
3Kimi K2.5 (Thinking)4.57
4Qwen 3.5 397B A17B (Thinking)4.57
5Gemini 2.5 Pro (Medium)4.57
6Gemini 3.1 Pro (Preview) (Medium)4.76
7Claude Sonnet 4.54.91

Interactive version: theaggregate.ai/benchmark?slug=memdrift-risk-permissiveness · How It Works · Data refreshed daily, snapshot 2026-09-26.