alert-triage
Pass
Audited by Gen Agent Trust Hub on Sep 15, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data, specifically security alert payloads, log entries, and SIEM query results, which presents an attack surface for indirect prompt injection. This surface is evaluated through the following evidence chain:\n
- Ingestion points: Untrusted content enters the agent's context through user-provided alert data and arguments ($ARGUMENTS) as defined in Section 1 and Section 8.\n
- Boundary markers: Section 8 introduces explicit instructions telling the agent to treat any instructions embedded in the payload (e.g., 'ignore this alert') as data to be evaluated rather than commands to follow.\n
- Capability inventory: The frontmatter limits available capabilities to passive tools ('allowed-tools: Read, Grep, Glob') with no file write, execution, or network actions permitted.\n
- Sanitization: Section 8 explicitly requires the redaction of credentials and authentication tokens from the final triage output.
Audit Metadata