🏃 Vendor Agents, Runners, and Build-Integration Trust Boundaries
Intro: Agents and runners are attractive because they automate a lot. They are dangerous for the same reason. Once deployed, they often gain filesystem, network, secret, or code visibility far beyond what the team consciously reviewed.
What this page includes
- trust-boundary review for agents and runners
- questions for vendor-hosted and self-hosted execution
- containment ideas when an agent or runner is suspected
- how to separate convenience from authority
Review questions
- what code will this component execute, and who controls that code?
- what network destinations can it reach?
- what secrets or service accounts are present at runtime?
- can it mutate artifacts, manifests, or production configuration?
- what telemetry exists if it behaves badly?
Safer patterns
- dedicate higher-trust runners for production paths;
- minimize mounted secrets and workspace reuse;
- quarantine suspicious runners quickly;
- keep vendor agents away from the most sensitive tenants or environments until proven necessary.
Related pages
Author attribution: Ivan Piskunov, 2026 - Educational and defensive-engineering use.