The Agent Skills Directory

[SAFE]: The skill implements a robust security model for managing sensitive risk policies. It explicitly instructs the agent to use the platform's authenticated connection and forbids the request of user credentials. Key safety features include:
Mandatory user confirmation for every project selection and policy modification.
Use of unique identifiers (slugs, IDs, URNs) instead of names or emails to prevent identity confusion.
Explicit refusal to perform self-removal or effective exclusion, directing users to administrative channels instead.
Implementation of idempotency keys and version checks to prevent conflicting updates.
Verification of state after write operations to ensure the intended changes were committed accurately.
[INDIRECT_PROMPT_INJECTION]: While the skill ingests data from external platform tools (e.g., list_projects, list_risk_policies) and performs write operations, it mitigates injection risks through multiple layers of verification:
Ingestion points: Data enters via list_projects, list_risk_policies, and get_risk_policy calls in SKILL.md.
Boundary markers: The instructions require the agent to stop and ask the user for confirmation at every critical step, such as choosing slugs or confirming audience replacements.
Capability inventory: The skill uses change_risk_policy_audience and update_risk_policy for modifications, all restricted to authenticated platform tools.
Sanitization: The agent is directed to use exact identifiers and URNs from trusted selections and to verify the expected_version before writing, reducing the risk of processing malicious metadata embedded in policy fields.

manage-risk-policy-audience