coze-workflow
Pass
Audited by Gen Agent Trust Hub on Sep 15, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill parses user-provided natural language specifications to structure and generate automation graphs, creating a surface for indirect prompt manipulation.
- Ingestion points: Natural language user-supplied process descriptions received by the agent during workflow orchestration as described in
SKILL.md. - Boundary markers: Absent; there are no explicit delimiter boundaries or instruction encapsulations enforced when mapping user requirements into code parameters.
- Capability inventory: Script execution capabilities via Bash are utilized to run the generated tool logic (
scripts/coze_yaml_builder.py). - Sanitization: Basic quote escaping is performed inside
scripts/coze_yaml_builder.pyvia_esc(), but the code compilation layer does not strictly validate or sanitize malicious sub-strings prior to script execution. - [DYNAMIC_EXECUTION]: The workflow creation architecture mandates that the agent dynamically compile a temporary Python script from templates and execute it locally using system tools to build the targeted ZIP package.
Audit Metadata