deepseek-usage

Warn

Audited by Gen Agent Trust Hub on Aug 14, 2026

Risk Level: MEDIUMCREDENTIALS_UNSAFEEXTERNAL_DOWNLOADSPROMPT_INJECTION
Full Analysis
  • [CREDENTIALS_UNSAFE]: The skill downloads and parses amount-2026-5.csv, which contains users' API keys in plain text. Although the instructions explicitly direct the agent to skip the index containing the key (p[4]), the presence of these credentials in the execution context poses a risk of accidental disclosure.
  • [EXTERNAL_DOWNLOADS]: The skill fetches the JSZip library from Cloudflare's well-known CDN (cdnjs.cloudflare.com) at runtime to handle file extraction.
  • [PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface by ingesting external CSV data from the DeepSeek platform. Ingestion point: platform.deepseek.com export API; Boundary markers: Absent for the raw CSV content; Capability inventory: execute_js and browser_use; Sanitization: The script uses manual string splitting and selective indexing (p[2], p[5], p[7]) to filter the data before it is presented to the agent.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Aug 14, 2026, 02:21 PM
Security Audit — agent-trust-hub — deepseek-usage