The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze external code, diffs, and logs, which creates a surface for indirect prompt injection. Attackers could potentially embed instructions within the candidate code to influence the agent's review verdict.
Ingestion points: The skill explicitly reads original materials, diffs, acceptance contracts, and logs provided as input for review.
Boundary markers: The instructions provide behavioral guidelines for independence but do not define technical delimiters to isolate the code being analyzed from the agent's instructions.
Capability inventory: The skill instructions mention the ability to run 'directional verification' when safe, suggesting the agent may have access to tools for code execution or test running.
Sanitization: No specific mechanisms for sanitizing or escaping the content from the processed files are described.
[NO_CODE]: The skill consists entirely of Markdown and YAML files for instruction and configuration. It does not include executable scripts, shell commands, or binaries, which reduces the potential for direct code-based attacks or unauthorized system modifications.

afk-slice-reviewer