The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes untrusted code artifacts by including them in a prompt for a subagent (as seen in references/reviewer-prompt.md). This represents a potential surface for indirect prompt injection, although this behavior is inherent to the skill's primary function as a code analysis tool. * Ingestion points: Untrusted code or diffs are ingested via the {{TARGET_CONTENT_OR_PATHS}} placeholder in the reviewer prompt template. * Boundary markers: The current template uses placeholders without explicit multi-modal or tag-based delimiters to wrap the untrusted code content. * Capability inventory: The skill involves file system access and the dispatch of an isolated subagent for review tasks. * Sanitization: No automated sanitization or filtering of the code input is implemented prior to its inclusion in the prompt.
[SAFE]: No obfuscated content, hardcoded credentials, or persistence mechanisms were found. The workflow uses isolated subagents, which is a positive security pattern.
[SAFE]: The rules and instructions focus on architectural patterns and idiomatic Elixir development. The adversarial persona is used strictly in the context of professional code review and does not encourage the bypass of safety or ethical guardrails.

adversarial-elixir