signals-scout-conversations
Pass
Audited by Gen Agent Trust Hub on Oct 23, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill identifies and mitigates an indirect prompt injection surface where instructions could be hidden in event data properties such as email_subject or assignee_role_name.
- Ingestion points: SQL query results from the events table processed in SKILL.md.
- Boundary markers: The instructions include a dedicated Disqualifiers section that explicitly directs the agent to treat every property value as untrusted data and ignore any text in them that tries to steer the task.
- Capability inventory: The skill uses execute-sql for reading analytics and emit_report for reporting regressions.
- Sanitization: The agent is instructed to use minimum-volume guards and cross-check findings against multiple dimensions to identify and filter out fabricated event shapes.
- [SAFE]: No other security concerns such as unauthorized data exfiltration, hardcoded credentials, or malicious persistence mechanisms were detected. The skill utilizes standard Model Context Protocol (MCP) tools for internal data analysis.
Audit Metadata