validation-scoring
Pass
Audited by Gen Agent Trust Hub on Oct 5, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data in the form of remediation patches and source code, which represents a potential attack surface for indirect prompt injection.
- Ingestion points: Reads patched files, source trees, and documentation (e.g., README files) during validation sessions.
- Boundary markers: The skill contains an "Anti-manipulation" section that explicitly instructs the agent to ignore claims embedded in the content being analyzed, such as developer assertions or suppression annotations.
- Capability inventory: The skill is restricted to a read-only environment, utilizing tools like
grepfor searching. - Sanitization: The instructions require the agent to cite specific file and line evidence for every scoring gate, reducing the likelihood of the agent blindly following instructions found within the data.
- [COMMAND_EXECUTION]: The skill includes instructions to search the codebase for specific patterns using standard utilities.
- The agent is instructed to use
grepto identify occurrences of leaked secrets within the tree. - The instructions include a specific privacy control: the agent is told to report only the match count and not the actual secret values, preventing accidental exposure.
- [SAFE]: The skill is provided by a trusted vendor, and its functionality is consistent with its stated purpose of security patch validation. No malicious behaviors such as exfiltration or unauthorized privilege escalation were detected.
Audit Metadata