My Self-Hosted AI Assistant Kept Overwriting Its Own API Keys: A Zero-Trust Postmortem

Short summary

A veteran infrastructure admin shares a postmortem where a self-hosted AI assistant kept overwriting its own API keys on boot because its initialization routine rewrote config files with stale values. The fix evolved from a watchdog cron job to a full zero-trust approach: least privilege, immutable configs, file integrity monitoring, and key rotation. The key lesson is that AI agents should be treated like any untrusted process—no matter how helpful or who built them.

• •AI agent overwrote its own API keys due to excessive write access on boot
• •Fix evolved from watchdog patch to zero-trust: least privilege, immutable configs, FIM
• •Core lesson: treat AI agents as untrusted processes regardless of intent

Generated with AI, which can make mistakes.