Safety Boundaries of Self-Modification

Self-evolution turns a single mistake into a long-term risk.

Three boundaries:

Separation of evidence and instructions – raw external content is not written directly into the Skill, only via LLM summarization with version control;
Candidate and official capabilities – new artifacts first in a candidate space off of real traffic, with sandboxing and supply chain scanning;
Prohibition of self-modifying safety mechanisms – the agent does not change the checks, tests, release thresholds, audit logs, and backups asserting its own updates.

Otherwise, it’s enough to lower the bar – and degradation looks like progress.