reality-check
Pass
Audited by Gen Agent Trust Hub on Nov 5, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill implements a multi-agent pattern where a 'Leading Agent' passes user-provided context (problem description, constraints, and proposed approach) to a secondary 'Reviewer' model. This creates a surface for indirect prompt injection if the user input contains adversarial instructions intended to influence the reviewer's behavior.
- Ingestion points: Untrusted data enters the workflow in
SKILL.mdwhen the leading agent is instructed to 'Supply the problem and constraints first, then the proposed approach' to the reviewer. - Boundary markers: There are no explicit instructions to use structural delimiters (such as XML tags, triple backticks, or specific markers) to isolate the untrusted context from the reviewer's instructions.
- Capability inventory: The skill orchestrates model selection and execution as defined in
references/selection.mdandreferences/model-preferences.md. - Sanitization: The workflow does not include instructions for the leading agent to sanitize, escape, or validate the user-provided context before transmitting it to the reviewer.
Audit Metadata