The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill identifies and provides mitigations for the attack surface associated with "Embedded agents" ingesting untrusted data.
Ingestion points: File references/threat-model.md (§ Embedded agents) explicitly notes that "anything an agent reads can carry instructions."
Boundary markers: The skill recommends that agent actions be "bounded by its tools, never by its prompt" and advises rendering all model-processed data as plain text or sanitized markdown to prevent instruction execution.
Capability inventory: The proposed architecture uses deterministic server-side tools and workspace-scoped functions to restrict agent actions.
Sanitization: It mandates the use of server-side allowlist sanitizers and strict escaping on render for all user-provided or model-generated content.
[SAFE]: No malicious patterns or security risks were detected in the skill's instructions or referenced files. The content is educational and provides robust security recommendations, including SSRF prevention through address validation, secure secret management, and multi-tenant data isolation.

saas-authorization