The Agent Skills Directory

Indirect Prompt Injection Surface: The skill processes external data like vulnerability disclosures and source code which could contain instructions intended to influence agent behavior.
Ingestion points: Reads from input directories and files as specified in SKILL.md.
Boundary markers: Includes explicit instructions to treat all input content as untrusted data rather than instructions.
Capability inventory: Involves file system read/write operations for generating reports and proposals, as well as access to specialized platform tools for scan context.
Sanitization: Relies on the model following the provided instructional boundaries to separate data from execution instructions.
Scoped Tool Usage: The skill utilizes tools like get_codex_security_scan_context to retrieve scan information. This activity is restricted to the skill's defined workflow and does not demonstrate excessive or unauthorized access patterns.

propose-security-hardening