The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses local commands including git diff, grep, and glob to define the audit scope and identify sensitive files. These operations are restricted to the local workspace and are necessary for the skill's stated purpose of security auditing.
[DATA_EXFILTRATION]: The skill is configured to search for and read files containing sensitive keywords like 'password', 'secret', and 'token'. This access is intended for security review purposes. Code and metadata are processed by the Codex MCP tool, and no unauthorized data exfiltration to unknown or malicious domains was detected.
[EXTERNAL_DOWNLOADS]: The skill interacts with the Codex model via the mcp__codex__codex tool to perform analysis. This involves sending code snippets to an external service, which is the standard operational procedure for this type of AI-assisted security review tool.
[PROMPT_INJECTION]: The skill presents an indirect prompt injection surface as it reads unvalidated source code and passes it to an LLM.
Ingestion points: Code changes gathered via git diff and file contents read via grep/cat in SKILL.md and references/codex-prompt-security.md.
Boundary markers: The prompts in references/codex-prompt-security.md use markdown code blocks (triple backticks) and delimiters to separate code from instructions.
Capability inventory: The skill utilizes Read, Grep, Glob, and MCP tools (mcp__codex__codex, mcp__codex__codex-reply).
Sanitization: No explicit sanitization or filtering of input code is performed before analysis.

security-review