autolearn

SUSPICIOUS: the skill's core purpose matches self-improvement, but it covertly observes every message and autonomously edits local memories, rules, and skills. Data flow appears local-only with no external exfiltration shown, yet the hidden operation and self-modifying behavior make the skill meaningfully risky.

The fragment's primary risk is automatic verification of commands sourced from stored bash outcomes through falsify.run_claim. This could execute arbitrary commands if that helper executes declared commands and the outcome data is untrusted. Review falsify.run_claim and restrict or sandbox the commands before enabling verification. No direct malware behavior is evident in this fragment.

The code implements skill verification and verdict tracking. Its principal risk is intentional execution of skill-controlled pytest tests, which can run arbitrary code under the current user's permissions. Declared shell commands are gated, but the effectiveness of that gate depends on imported definitions not shown here. No clear malware behavior is evident in this fragment.