The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests data from various external sources, including third-party skill registries and general ecosystem research.
Ingestion points: SKILL.md identifies multiple registries (e.g., Agent Garden, skills.sh) and the core/research.skill as sources for new capabilities.
Boundary markers: The skill references GABBE_AUTONOMY gating and validation checks as boundaries for processing untrusted content.
Capability inventory: The skill can create new git branches (evolve/*), write to AUDIT_LOG.md, and execute validation tools (validate_skills).
Sanitization: Implements frontmatter parsing, slug/path-traversal checks, and scans for secrets, egress patterns, and executable content.
[DYNAMIC_EXECUTION]: The skill is architected to dynamically update the agent's logic by adopting new skills and tools discovered at runtime.
Evidence includes the Decision loop which allows for the automatic adoption of changes deemed cheap and reversible.
Mitigations are documented within the skill: mandatory validation for external content, human approval for all external code, and the requirement for test suite success before merging changes.

update-scan