The Agent Skills Directory

[Indirect Prompt Injection Surface]: The skill ingests external vulnerability disclosures and incident reports (ingestion points in SKILL.md). It directs the agent to treat this content strictly as untrusted data, functioning as a boundary marker. The skill is limited to reading local files and writing report artifacts (capability inventory in Workflow Step 8), with sanitization enforced by instructions to ignore embedded instructions. This approach reduces the likelihood of external data influencing agent behavior.
[Local File Operations]: The skill creates structured JSON and Markdown artifacts within the project's directory. These operations are restricted to generating derived design products and do not include unauthorized data transmission or persistence mechanisms. The use of repository-relative paths in the proposal-format.md prevents the leakage of local filesystem information.
[Platform Tool Integration]: The skill utilizes platform-specific tools, such as get_codex_security_scan_context, for accessing security scan contexts. This integration ensures the analysis is based on verified evidence provided by the secure execution environment rather than potentially tampered local files.

propose-security-hardening