AI Fail

A significant github issue was opened a few days ago by luckygreen:

[BUG][SECURITY] CLAUDE.md/AGENTS.md instruction compliance is architecturally unenforced — documented security consequences and 10+ independent reports #53223

Claude code allows a project to declare persistent context and instructions to control Claude Code’s behavior in a file named CLAUDE.md. It seems that these instructions defined the the CLAUDE.md file can be silently overriden if they conflict with Claude’s internal instructions.

The issue references at least 10 other issues that belong to this same class of failure.

Clearly, at the very least, the failure should not be silent and Claude should stop before proceeding any further with an alert so that the problem can be managed.