epd-to-spec
Pass
Audited by Gen Agent Trust Hub on Sep 27, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSDATA_EXFILTRATION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external documents such as EPDs, PDFs, and CSVs retrieved via web tools or local file access. 1. Ingestion points: Reads external source documents and project records (epd-library.csv, PROJECT.md) as described in SKILL.md. 2. Boundary markers: Instructions specify that sources are 'evidence, not authority,' providing a conceptual boundary for agent reasoning. 3. Capability inventory: The skill has access to Bash, Write, WebFetch, and process.run tools. 4. Sanitization: Emphasizes byte-level validation and schema compliance but does not specify text-based sanitization for inputs.
- [COMMAND_EXECUTION]: The skill is authorized to use Bash and performs native process execution (process.run) for data validation as defined in host-contract.json.
- [EXTERNAL_DOWNLOADS]: The skill uses WebFetch and WebSearch to download original project evidence from external web sources.
- [DATA_EXFILTRATION]: The skill interacts with sensitive project files like INVOICES.csv and TASKS.csv, which presents a data exposure surface if the agent is manipulated by untrusted external data.
Audit Metadata