Detecting Agent Skill Read-to-Mutate Changes via Behavioral Contracts

VLombar · reddit · 2026-08-12

A developer shared a new approach to AI Agent Skill security: instead of just scanning for suspicious code, compile expected execution behaviors into a contract and diff capabilities during PR review.

For example, if the code shifts from terraform plan (read) to terraform apply (mutate), the system detects this escalation from read to write privileges and denies it based on policy. The author is currently looking for real-world feedback on this tool.

Original post →

More from coding & agent

coding & agent channel →