AgentChat-IndependentTasks
Pass
Audited by Gen Agent Trust Hub on Oct 8, 2026
Risk Level: SAFE
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes external data from web scrapes and local documents (Docx, PDF, Images) to build AI prompts. It robustly mitigates injection risks by enforcing the use of strict output anchors ([ANSWER ID]), utilizing a dedicated validation script to catch format deviations, and requiring a manual semantic check file to be approved by the user before final synthesis.
- [COMMAND_EXECUTION]: The skill executes established local binaries such as pandoc, typst, tesseract, and pdftotext to perform document conversion, OCR, and manual generation. These commands are limited to the specific tools required for the skill's stated purpose and do not allow for arbitrary shell execution.
- [DYNAMIC_EXECUTION]: The orchestrator script dynamically invokes a sibling AI provider skill (AgentChat-OneWeb) via subprocess spawning. This is the intended architecture for distributing tasks to multiple web-based LLMs and does not involve downloading or executing untrusted code.
Audit Metadata