claims-drafting
Pass
Audited by Gen Agent Trust Hub on Nov 10, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill incorporates content from multiple local files into prompts for a spawned reviewer agent, creating a surface for indirect prompt injection.
- Ingestion points: Content is read from
patent/INVENTION_DISCLOSURE.md,patent/PRIOR_ART_REPORT.md, andpatent/NOVELTY_ASSESSMENT.mdas described in Step 5. - Boundary markers: The prompt templates used to spawn the reviewer agent do not employ delimiters or specific instructions to ignore embedded commands within the ingested file data.
- Capability inventory: The agent has access to powerful tools including
Bash(*),WebSearch,WebFetch, and file system modification tools (Write,Edit). - Sanitization: There is no evidence of sanitization or validation of the external file content before it is interpolated into the prompt strings.
- [DYNAMIC_EXECUTION]: The skill employs dynamic agent orchestration to facilitate the patent review process.
- The workflow in Step 5 and 6 dynamically generates instructions and contexts to create new agent instances (
spawn_agent) and interact with them (send_input) at runtime.
Audit Metadata