multi-ai-research

Warn

Audited by Gen Agent Trust Hub on Apr 9, 2026

Risk Level: MEDIUMEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill requires the installation of a third-party CLI tool, @jackwener/opencli, and an associated browser bridge extension from an unverified GitHub repository. This dependency is central to the skill's ability to communicate with external AI models.
  • [COMMAND_EXECUTION]: The agent is instructed to execute multiple shell commands, including 'opencli' for research tasks and Python/SQLite scripts for data health checks. These commands involve environment variable manipulation and interactions with the local file system.
  • [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it ingests data from external AI services that process live web content and X (Twitter) feeds. Malicious instructions hidden in those external data sources could potentially influence the research results or the agent's behavior.
  • Ingestion points: Data retrieved via 'opencli grok' and 'opencli gemini' from the web/social media (SKILL.md, Phase 3).
  • Boundary markers: The skill uses internal templates and structural requirements (Phase 2) but lacks strict cryptographic or non-textual boundary markers for external data.
  • Capability inventory: Execution of bash commands, file system access via Python/SQLite, and dispatching multiple internal sub-agents.
  • Sanitization: Includes a 'Phase 5' arbitration logic that cross-validates findings and checks for 'potential hallucinations', which serves as a functional mitigation against data poisoning.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Apr 9, 2026, 10:50 PM