perfect-company-list
Pass
Audited by Gen Agent Trust Hub on Oct 1, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONDATA_EXFILTRATION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted company metadata (name, industry, description) from external sources or CSV files into AI prompts for qualification. Ingestion points:
scripts/jev_judge.pyreads data from input CSV files;clay-workflow.mdprocesses data from enrichment tools. Boundary markers: The prompt uses specific indexing instructions (e.g.,Answer only about companies[{i}]) to attempt to isolate data, though this is a weak boundary. Capability inventory: The skill performs network operations to external APIs and file writing to local CSVs. Sanitization: Input descriptions are truncated to 300 characters inscripts/jev_judge.py, but no specific character escaping or instruction filtering is applied. - [COMMAND_EXECUTION]: The skill provides a standalone Python script
scripts/jev_judge.pyand instructions for using theclayCLI, which involve local execution of commands to process data batches. - [DATA_EXFILTRATION]: The script and workflows make network requests to
api.typesafe.aito send company metadata for AI judging. This represents a network dependency on a non-whitelisted domain.
Audit Metadata