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_email and financial transactions.
  • Sanitization: Defensive measures such as PII redaction, schema validation, and human-in-the-loop authorization are recommended.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 9, 2026, 07:33 PM
Security Audit — agent-trust-hub — guardrails-and-safety