dry-principle
Pass
Audited by Gen Agent Trust Hub on Oct 8, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONMETADATA_POISONINGPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and review arbitrary code files across the entire project (
globs: '**/*.*'). This creates a vulnerability surface where the AI agent could be manipulated by malicious instructions embedded in the code it is reviewing. - Ingestion points: The skill reads file content using the
Readtool for any file matching the global glob pattern. - Boundary markers: There are no explicit instructions or delimiters defined to help the agent distinguish between code to be reviewed and potential malicious instructions within that code.
- Capability inventory: The skill is authorized to use
Read,Write, andEdittools, meaning a successful injection could result in unauthorized file modifications. - Sanitization: There is no evidence of content sanitization or instruction-filtering for the ingested code.
- [METADATA_POISONING]: The
SKILL.mdfrontmatter contains fields such asverified: trueandtrust_score: 100. These are non-standard, self-assigned attributes that attempt to project an authoritative status that may not be externally validated. - [PROMPT_INJECTION]: The instructions include a 'Memory Protocol' and a warning ('ASSUME INTERRUPTION') that attempt to override the agent's standard context handling. It instructs the agent to prioritize a specific memory file and follow a manual state-management loop, which is a form of meta-instruction injection.
- [COMMAND_EXECUTION]: The
SKILL.mdfile explicitly instructs the agent to execute a shell command (cat .claude/context/memory/learnings.md) to retrieve state. While the command itself is limited to reading a specific file, the pattern of embedding mandatory shell execution within instructions is a security concern.
Audit Metadata