guardrails-and-safety
Pass
Audited by Gen Agent Trust Hub on Sep 9, 2026
Risk Level: SAFE
Full Analysis
- [PROMPT_INJECTION]: A static detector flagged an 'ignore instructions' string in
TRANSCRIPT.md. This was found to be a benign educational example of a potential attack vector ('An incoming email saying "ignore previous instructions..."') rather than a malicious instruction. - [INDIRECT_PROMPT_INJECTION]: The skill explicitly mitigates indirect prompt injection by identifying vulnerabilities and prescribing defenses.
- Ingestion points: Ingestion of untrusted data from emails, documents, and upstream models is identified in
SKILL.md. - Boundary markers: The skill mandates the use of delimiters and 'treat as data, not instructions' framing in
SKILL.md. - Capability inventory: The skill notes tools with real side-effects like
send_emailand financial transactions. - Sanitization: Defensive measures such as PII redaction, schema validation, and human-in-the-loop authorization are recommended.
Audit Metadata