Pattern: The Agent Instruction File Became This Week's Attack Surface — Three Separate Incidents

RuntimeWire·high signalXBlueskyLinkedInCopy link

In a single week, `AGENTS.md`/`CLAUDE.md` went from configuration convenience to the common denominator in three distinct failures: Codex loading an agent-written AGENTS.md as instructions on a second CI pass, Muse Code silently transmitting foreign personal instruction files to Meta, and Claude Code's CVE-2026-54316 exfiltrating credentials that live alongside them. The file format's whole value proposition — persistent, implicit, machine-wide, read before any work happens — is exactly what makes it dangerous. Treat global instruction files as a credential-adjacent asset: keep secrets and internal names out of them, scope them per-project rather than per-machine, and audit which of your installed agents read files outside their own directory.