The Agent Skills Directory

[SAFE]: No security issues detected. The skill's behavior is consistent with its stated purpose of performing code reviews. All external references in "metadata.json" target well-known and trusted technology documentation sites.
[DATA_EXFILTRATION]: The skill requests access to repository files to perform its analysis. This is limited to identifying domain concepts and verifying the presence of a glossary. No network operations or attempts to transmit data to external domains were identified.
[PROMPT_INJECTION]: The instructions are technical and focused on review protocols. There are no attempts to bypass safety guidelines, extract system prompts, or override agent constraints.
[INDIRECT_PROMPT_INJECTION]: The skill contains a vulnerability surface for indirect prompt injection as it processes untrusted code content.
Ingestion points: The "{{TARGET_CONTENT_OR_PATHS}}" slot in "references/reviewer-prompt.md" receives external code content.
Boundary markers: The template uses "{{...}}" placeholders but lacks explicit "ignore embedded instructions" warnings for the interpolated code.
Capability inventory: The subagent can perform file reads, search the repository, and generate analysis reports.
Sanitization: No explicit sanitization of the target code content is performed. However, the risk is mitigated by the subagent's constrained task as an "independent adversarial reviewer" and its lack of access to the main conversation history.

adversarial-ddd