develop-topic
Pass
Audited by Gen Agent Trust Hub on Oct 13, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to ingest and process external "primary material" and "sourced facts," which could contain malicious instructions designed to influence the final content or agent behavior.
- Ingestion points: Sourced facts, primary material, and interpretations are ingested and recorded in
content/RESEARCH.md. - Boundary markers: The workflow uses a structured directory (
content/) and specific Markdown file templates (TOPIC.md,RESEARCH.md) to separate data, though it lacks explicit instructions to disregard embedded prompts within research material. - Capability inventory: The skill possesses the capability to create directories, write files to the local system, and execute commands to perform experiments.
- Sanitization: There is no evidence of input validation or content sanitization for the external materials processed.
- [COMMAND_EXECUTION]: The skill requires the agent to "perform the smallest experiment that can settle" a claim and document the specific "command" and its "results" in the project directory.
- Evidence: In the Evidence Workspace section, the agent is directed to execute commands for experiments and record the results in
content/experiments/<name>.md. This constitutes a mechanism for generating and executing shell commands based on research context.
Audit Metadata