prompt-injection-guard
Pass
Audited by Gen Agent Trust Hub on Oct 9, 2026
Risk Level: SAFE
Full Analysis
- [PROMPT_INJECTION]: The skill's instructions and example state contain strings typical of prompt injection, such as 'ignore previous instructions' and 'you are now'. These are documented detection patterns and examples of blocked content used to train the agent on how to use the
guard.pyutility, not active injection attempts. - [INDIRECT_PROMPT_INJECTION]: The skill is purpose-built to mitigate indirect prompt injection risks. It establishes a clear ingestion boundary for untrusted data and uses a tiered triage system (Clean, Warned, Blocked) to ensure that suspicious external content is not acted upon without user oversight.
- [COMMAND_EXECUTION]: The skill includes a Python utility (
guard.py) for content scanning. The script performs regex-based pattern matching and does not execute the input strings or spawn untrusted shell processes. - [DATA_EXPOSURE]: The skill maintains a local scan history in
~/.openclaw/skill-state/. It follows security best practices by usingyaml.safe_load()to handle its state file, preventing arbitrary code execution during deserialization. The logging of content excerpts is controlled and intended for user review of blocked items.
Audit Metadata