The original title is "Running Untrusted Code Safely: A Field Guide for AI and CI Pipelines"

Short summary

Untrusted code in CI pipelines (via AI reviewers, fork PRs, or bots) demands isolation beyond sandboxing: keep secrets out of runner environments, scope credentials narrowly, disable plugin loading from untrusted repos, and deny network egress by default. Assume sandbox failures; design so they are contained, not catastrophic.

• •Never execute untrusted code directly; secure the automation and tools that run around it instead
• •Isolate credentials completely: no platform tokens, database passwords, or signing keys in the container
• •Disable plugin and config loading from untrusted repos; apply least privilege at every layer

Generated with AI, which can make mistakes.