campaign-angle-finder
Pass
Audited by Gen Agent Trust Hub on Sep 29, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data including company names, domains, and evidence descriptions from external rows provided in a table or CSV (SKILL.md, Declared inputs).
- Ingestion points: External data enters the context in Step 2 and is interpolated into the model prompt in Step 3.
- Boundary markers: The instructions describe a structured prompt assembly that separates the system offer, examples, and row data into discrete blocks to assist the model in distinguishing data from instructions.
- Capability inventory: The skill is restricted to writing record columns and does not have capabilities for network operations, shell command execution, or file system access.
- Sanitization: Rule 5 implements a verification step that requires the model to provide a quote from the evidence; the process then asserts that this quote is a verbatim substring of the input, which prevents the model from fulfilling instructions that would result in invented or irrelevant output.
Audit Metadata