The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze untrusted code provided by users for security auditing, which creates a significant attack surface.
Ingestion points: The skill is explicitly triggered to review files, diffs, and code snippets provided in the agent's context (SKILL.md).
Boundary markers: The skill defines a structured review process using specific confidence levels (HIGH, MEDIUM, LOW) and a required output format to maintain control over the agent's behavior (SKILL.md).
Capability inventory: The agent is granted access to powerful system tools including Bash, Task, Read, Grep, and Glob, allowing it to perform deep analysis of the provided code (SKILL.md).
Sanitization: Instructions include specific mandates to redact secrets and use masked placeholders to prevent accidental exposure (SKILL.md).
[DYNAMIC_EXECUTION]: The reference guides include examples of dangerous dynamic execution patterns (e.g., eval(), exec(), Pickle.loads()) as targets for detection during the review process (SKILL.md, references/deserialization.md, languages/python.md).
[COMMAND_EXECUTION]: The supply chain reference file contains examples of reverse shell commands and OS command injection vectors to assist the agent in identifying these patterns in untrusted packages (references/supply-chain.md, references/injection.md).
[OBFUSCATION]: The skill documentation includes examples of Base64-encoded strings that decode to system commands (e.g., aW1wb3J0IG9z to import os) to demonstrate how malware might hide its intent (references/supply-chain.md).

security-review