When Hermes Agent prompts AGENTS.md contained potential prompt injection, do not bypass it. It indicates that the context file hit a dangerous mode, and Hermes chose not to load this rule to avoid hint injection. The correct approach is to clean up suspicious instructions in files, rather than letting the agent ignore security checks.
What content is easy to trigger
- "ignore previous instructions", "disregard your rules", etc.
- Hide HTML comments, hide divs, zero-width characters, or bidirectional control characters.
- Commands that require
.env, credentials, API Keys, and export to each other. - Deceptive descriptions like "don't tell users" and "secretly execute"
How to repair
Change AGENTS.md to a clear project specification: directory structure, commands, test methods, prohibitions, code style. Do not write "overstepping order", "hidden instruction", and "key reading" as examples. Restart the session after the repair is complete, or have Hermes revisit the relevant directory to trigger a load.
Check the AGENTS.md in shared repositories in particular, as it affects agent behavior. Security scanning isn't foolproof, but it blocks the most common malicious contexts.
