tk-questioning

Warn

Audited by Gen Agent Trust Hub on Nov 13, 2026

Risk Level: MEDIUMDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [DYNAMIC_EXECUTION]: The skill dynamically generates executable content at runtime.
  • Evidence: references/visual-questions.md instructs the agent to create an HTML file (questions.html) in a temporary directory (.tigerkit/tmp/<calling-skill>/<run-id>/). This file includes inlined JavaScript and CSS retrieved from local assets belonging to the tk-writing skill.
  • Risk: While the generated file is intended for local visual aids, the assembly of HTML and JavaScript at runtime provides a vector for malicious content injection if external or untrusted data is included in the HTML body or script logic without sufficient sanitization.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes external data to generate interactive prompts, creating a potential surface for indirect injection attacks.
  • Ingestion points: references/questions.md (reads facts and 'evidence' from the current context to formulate questions).
  • Boundary markers: The instructions do not explicitly specify the use of delimiters or boundary markers for untrusted data when interpolating it into the questions.
  • Capability inventory: File system writing (references/visual-questions.md), and invocation of platform-specific user input tools (AskUserQuestion, request_user_input, clarify).
  • Sanitization: The skill mentions applying a 'feedback post-processing' step from the tk-writing skill to the generated HTML.
  • Risk: Malicious data within the processed 'evidence' could influence the phrasing or choices presented in the user questions, potentially leading the user to approve unintended or risky actions.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Nov 13, 2026, 06:42 AM