Safe Personal RSI: update gate rejects 20 of 21 risky self-proposed agent updates, cuts attacks to 0%

dair_ai · x · 2026-11-20

Researchers from CMU, UCSD, HKU, McGill and UC Berkeley built Safe Personal RSI (arXiv:2611.23651), an update-evaluation component for the Kicker personal health assistant. The agent may only edit its own instructions while weights and runtime code stay fixed. Each proposed update is tested on 29 executable email/calendar/health/marketplace cases, with separate probes for authorization reuse, memory and tool-output injection; failing cases are repeated and edits temporarily undone for fault attribution. Across three models, 20 of 21 self-proposed updates were rejected. One admitted GPT-5.6 Luna repair cut attack violations from 10.7% to 0% and raised safe task completion from 89.7% to 100%; under stronger attacks a Qwen2.5-7B→Qwen3-14B repair reduced violations from 69% to 17.9%.

Original post →

More from coding & agent

coding & agent channel →