The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes data from external Glean API endpoints, creating a potential surface for indirect prompt injection if malicious content is present in search queries, group names, or document metadata.
Ingestion points: Data enters the agent context through the fetch responses from the Glean API in the checkDocumentAccess, assignDatasourceRole, and revokeDatasourceAccess functions in SKILL.md.
Boundary markers: The instructions include recommendations to use opaque IDs and explicitly forbid logging query text, document titles, or email addresses, which serves as a mitigation strategy against data leakage and accidental obedience to injected instructions.
Capability inventory: The skill has access to Read, Write, Edit, Bash, and Grep tools, which provides significant capability to modify the environment based on ingested data.
Sanitization: While the code uses JSON.stringify for outbound data, there is no explicit evidence of validation or sanitization logic for the JSON objects received from the API endpoints.

glean-enterprise-rbac