meta-cognition-parallel

Pass

Audited by Gen Agent Trust Hub on Jun 25, 2026

Risk Level: SAFEPROMPT_INJECTION
Full Analysis
  • [PROMPT_INJECTION]: The skill contains an indirect prompt injection surface by design, as it interpolates raw user input into the context of sub-agent tasks.
  • Ingestion points: Untrusted data enters the context through the $ARGUMENTS variable in the SKILL.md parsing logic.
  • Boundary markers: The skill uses a simple markdown header (## User Query) to separate instructions from user data, which can be bypassed by adversarial input.
  • Capability inventory: The skill triggers general-purpose sub-agents using the Task primitive; it does not explicitly invoke dangerous capabilities like arbitrary shell execution or network requests in the primary file.
  • Sanitization: No input sanitization, escaping, or schema validation is applied to $ARGUMENTS before it is added to the sub-agent prompts.
  • [SAFE]: The skill attempts to access configuration files in a parent directory (../../agents/). This is interpreted as a standard file-system operation for a multi-component agent setup and does not target sensitive system credentials or user data.
Audit Metadata
Risk Level
SAFE
Analyzed
Jun 25, 2026, 05:48 PM
Security Audit — agent-trust-hub — meta-cognition-parallel