check-agent-access

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (medium risk: 0.30). At runtime the workflow iterates over agents in scope and reads each agent’s live spec via semantic_studio (Step 3), then extracts dependency FQNs found in instructions, which can include outsider-authored free text embedded in the agent specification.

MEDIUM W013: Attempt to modify system services in skill instructions.

• Attempt to modify system services in skill instructions detected (high risk: 1.00). The skill explicitly generates and can execute privileged GRANT statements (modifying Snowflake RBAC) and even suggests switching to SYSADMIN/ACCOUNTADMIN to apply them, which changes the system's access state and requires elevated privileges.