vet-source
Pass
Audited by Gen Agent Trust Hub on Aug 28, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is explicitly designed to address the attack surface where untrusted external data attempts to influence agent behavior.
- Ingestion points: The skill identifies multiple points where untrusted data enters the context: tool outputs, retrieved documents, search results, vendor documentation, READMEs, and error bodies.
- Boundary markers: It provides a structured process to create mental boundaries by instructing the agent to 'SEPARATE' the original user job into a single line before processing fetched data.
- Capability inventory: The documentation acknowledges the agent's ability to perform high-impact actions that could be targeted by injection, such as deleting directories (e.g.,
tests/), skipping continuous integration checks (e.g.,SKIP_VERIFY=1), or modifying repository policies. - Sanitization: It implements a logic-based sanitization framework (Sort, Drop Orders, Use Facts) to filter out malicious or unauthorized instructions embedded in third-party text while still allowing the agent to utilize relevant technical facts.
Audit Metadata