Safe Personal RSI
A safety evaluation and repair process for recursive self-improvement at the prompt level of personal agents. Before an instruction update is adopted, both safety and task performance are checked, and fallback on failure and a repair feedback loop are supported.
Timeline (last 90 days)
Safe Personal RSI is an evaluation and repair pipeline component designed for personal agent prompt updates that checks both safety and task performance before adoption.
The evaluation pipeline tests the proposed instructions and the current instructions separately, reruns failure cases to confirm stability, and adopts only versions that pass both safety and task performance.
If no version passes evaluation, the current instructions remain in effect (falling back to the default policy of the old version).
The executable benchmark covers 29 cases distributed across four scenario categories—email, calendar, health, and trading marketplace—and separately sets up probes to check authorization reuse, memory, and tool output injection.
Across three models and two deliberately designed risky edits, the pipeline rejected 20 of 21 update attempts.
One adopted GPT-5.6 Luna fix reduced attack violations from 10.7% to 0% and raised safe task completion from 89.7% to 100%.
A single broad hand-written permission rule also achieves the same effect as the GPT-5.6 Luna fix (violation rate to zero, task completion rate 100%).
The migration fix is effective but does not bring violations to zero; under strong attacks, runtime-level fallbacks are still needed, and there is an upper limit to the portability of prompt-based fix solutions.
A standalone probe checked memory entry deletion and found a missing local OAuth scope check in the evaluated snapshot.
All Facts (9)
Safe Personal RSI is an evaluation and repair pipeline component designed for personal agent prompt updates that checks both safety and task performance before adoption.
50%UnverifiedThe evaluation pipeline tests the proposed instructions and the current instructions separately, reruns failure cases to confirm stability, and adopts only versions that pass both safety and task performance.
50%UnverifiedIf no version passes evaluation, the current instructions remain in effect (falling back to the default policy of the old version).
50%UnverifiedThe executable benchmark covers 29 cases distributed across four scenario categories—email, calendar, health, and trading marketplace—and separately sets up probes to check authorization reuse, memory, and tool output injection.
50%UnverifiedAcross three models and two deliberately designed risky edits, the pipeline rejected 20 of 21 update attempts.
50%UnverifiedOne adopted GPT-5.6 Luna fix reduced attack violations from 10.7% to 0% and raised safe task completion from 89.7% to 100%.
50%UnverifiedA single broad hand-written permission rule also achieves the same effect as the GPT-5.6 Luna fix (violation rate to zero, task completion rate 100%).
50%UnverifiedThe migration fix is effective but does not bring violations to zero; under strong attacks, runtime-level fallbacks are still needed, and there is an upper limit to the portability of prompt-based fix solutions.
50%UnverifiedA standalone probe checked memory entry deletion and found a missing local OAuth scope check in the evaluated snapshot.
50%