WeaknessASC-2026-0016Major weakness
Poisoned agent configuration and instruction files
HighMitigated
Coding agents trust files in the repository: rules files, AGENTS.md, CLAUDE.md, .cursor and .claude settings, hooks and editor tasks. A malicious commit or dependency that writes them plants instructions or runs code the next time someone opens the project.
How to detect it
- Agent instruction or settings files changed by a dependency, a bot or an outside contributor
- Hooks or tasks that run on session start or folder open
- Hidden Unicode or very long lines in rules files
How to resolve it
- Review agent configuration files like CI configuration, with CODEOWNERS on .claude/, .cursor/, .vscode/, AGENTS.md and rules files
- Keep workspace trust prompts on and do not auto-run hooks in untrusted repositories
- Scan these files in CI for hidden characters, network commands and hooks
- Install dependencies with lifecycle scripts disabled so packages cannot write them
How to investigate
- Inspect git history of agent configuration files for unexpected authors and recent changes
- Check which machines opened the project after the change and what the hooks ran
- Rotate credentials available on those machines