← News

Agents & MCPSources: arXiv · 19 Nov 2026

Kicker study tests a safety gate for self-edited assistant prompts

Researchers behind the Kicker health assistant describe an update check that reviews proposed instruction changes for safety and task performance before any change is adopted.

In plain words

The team built a checkpoint that reviews changes to an assistant's written instructions before those changes take effect. Model weights and code stay fixed; only the instructions can change. If a proposed update fails testing, the assistant keeps its current instructions.

What was announced

The system, called Safe Personal RSI, limits recursive self-improvement to prompt updates while model weights and runtime code remain fixed. According to the paper, the procedure tests proposed and current instructions side by side, repeats failing cases, temporarily reverts changes to guide repairs, and checks both safety and task performance before adoption. When no candidate version passes, the existing instructions remain in use.

The accompanying benchmark spans 29 email, calendar, health, and marketplace cases, plus separate probes for authorization reuse, memory, and tool-output injection. Across three models and two deliberately risky edits, the procedure rejected 20 of 21 update attempts. In repair confirmation, the authors report that an accepted GPT-5.6 Luna repair cut attack violations from 10.7% to 0% and raised safe task completion from 89.7% to 100%; a broad handwritten authority rule matched those results. Under stronger attacks, a repair developed on Qwen2.5-7B-Instruct transferred to Qwen3-14B, lowering violations from 69.0% to 17.9% and lifting safe completion from 19.5% to 64.4%. Runtime probes also flagged missing local OAuth scope checks in the evaluated snapshot.

What it does not show

  • Evidence is limited to the paper's abstract; full methods, baselines, and peer review status are not visible in the supplied passages.
  • Reported rejection rates and repair gains come from the authors' own evaluation and have not been independently confirmed.
  • The OAuth scope-check finding applies to the evaluated snapshot and may not reflect the current software.

For security and IT teams

Scope
The benchmark covers email, calendar, health, and marketplace cases with separate probes for authorization reuse, memory, and tool-output injection, per the paper's abstract.

Sources

Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 20 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.