Safe Personal RSI: Evaluating and repairing personal agent updates
Safe Personal RSI: Evaluating and Repairing Personal-Agent Updates
The paper proposes Safe Personal RSI, an update evaluation component developed for the personal health assistant Kicker, which restricts recursive self-improvement to the prompt update level, with model weights and runtime code fixed; updates are adopted only if they pass safety and task performance checks, otherwise the current instructions are retained. The executable benchmark covers 29 cases across email, calendar, health, and a trading marketplace, plus probes for authorization reuse, memory, and tool output injection. Across three models and two deliberately high-risk edits, the pipeline rejected 20 of 21 update attempts; an adopted GPT-5.6 Luna fix reduced the attack violation rate from 10.7% to 0% and raised the safe task completion rate from 89.7% to 100%. Under a stronger attack, a fix migrated from Qwen2.5-7B-Instruct to Qwen3-14B reduced the violation rate from 69.0% to 17.9% and raised the safe task completion rate from 19.5% to 64.4%. Independent runtime probes also check memory entry deletion and found a missing local OAuth scope check in the evaluation snapshot.