design-draft
Pass
Audited by Gen Agent Trust Hub on Oct 31, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill reads Project Requirements Documents (PRDs) from an external directory (
../refinery-pm-os/outputs/prds/approved/) to drive its multi-step workflow. These files are processed to generate code and documentation, creating a surface for indirect prompt injection where malicious content in a PRD could influence the agent's behavior. - Ingestion points: Reads markdown files from
../refinery-pm-os/outputs/prds/approved/andcontext-library/prds/(SKILL.md, Step 0 and 1). - Boundary markers: None specified for the external data ingestion.
- Capability inventory: The skill can write files to multiple
outputs/subdirectories and execute a shell script (./scripts/uxd-approve.zsh). - Sanitization: No sanitization or validation of the PRD content is described before use in prompts.
- [DYNAMIC_EXECUTION]: The skill dynamically generates and modifies executable code at runtime.
- Evidence: Step 6 generates HTML files in
outputs/screens/with embedded JavaScriptonclickhandlers for navigation. Step 8b instructs the agent to programmatically updateoutputs/screens/manifest.jswith new data structures. - [COMMAND_EXECUTION]: The skill executes a local shell script to finalize the workflow.
- Evidence: Step 9 executes
./scripts/uxd-approve.zsh [FEATURE-ID]. TheFEATURE-IDis derived from external file names or user input.
Audit Metadata