design-study
Pass
Audited by Gen Agent Trust Hub on Oct 9, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data such as research protocols, drafts, slides, and tables during the 'Reconstruct the study' phase. This represents a potential attack surface where malicious instructions could be embedded in the clinical data to influence the agent's behavior.
- Ingestion points: Phase 1 in
SKILL.mdidentifies protocols, drafts, slides, tables, and notes as primary inputs. - Boundary markers: The skill does not define specific delimiters or instructions to ignore embedded prompts within these documents.
- Capability inventory: The skill uses a local Python script
scripts/adjustment_set_helper.pyand a bash verification scriptscripts/adjustment_set_challenge/verify.sh. It also writes decision notes. - Sanitization: No explicit sanitization or filtering of input data is documented.
- [COMMAND_EXECUTION]: The skill utilizes a local Python utility
scripts/adjustment_set_helper.pyto assist with confounder selection in causal diagrams (DAGs). While this involves script execution, the utility uses standard libraries and processes data locally without network access or high-privilege operations.
Audit Metadata