Dev trusts coding agents with production secrets, given isolation and least privilege

rseroter · x · 2026-09-10

AI infra engineer Jake Gold explains why he now hands his coding agents production secrets—Tailscale keys, SSH private keys, Anthropic/Codex OAuth tokens—stored in a vault so agents use them without seeing values. His setup hits Simon Willison's 'lethal trifecta', but he argues security is a productivity-risk tradeoff: with production access, agents can investigate outages, deploy fixes, and verify them autonomously. He compares it to onboarding new teammates and says frontier models' improved injection resistance makes this viable—something he wouldn't do with older models. Preconditions: isolate agents and apply least privilege.

Original post →

More from coding & agent

coding & agent channel →