research-integration

Warn

Audited by Gen Agent Trust Hub on May 20, 2026

Risk Level: MEDIUMEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [REMOTE_CODE_EXECUTION]: The skill generates a standalone Python script (test-api.py) based on research findings derived from untrusted external documentation and vendor repositories. It also performs a syntax check on the generated file using the python3 compiler, which executes the script logic at a basic level.\n- [EXTERNAL_DOWNLOADS]: The skill instructs subagents to clone git repositories and install software packages using pip or npm from external sources to analyze vendor-specific SDKs and schema definitions during the research process.\n- [COMMAND_EXECUTION]: Research subagents are authorized to execute Python analysis scripts against raw data artifacts, and the orchestrator executes shell commands to verify generated code and manage directory structures.\n- [PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface by ingesting and synthesizing data from arbitrary documentation URLs, vendor portals, and code repositories.\n
  • Ingestion points: Documentation URLs, git repositories, and SDK packages fetched during subagent research tasks (Phase 2).\n
  • Boundary markers: The skill does not provide instructions to use delimiters or ignore-embedded-instruction warnings when subagents process external documentation.\n
  • Capability inventory: Subagents can perform network fetches, clone repos, install packages, and execute Python scripts, which can be leveraged if a subagent is influenced by malicious instructions in a processed document.\n
  • Sanitization: While the skill advises anonymizing sample data (IPs, names), it lacks generic sanitization or validation for the content of fetched documentation.
Audit Metadata
Risk Level
MEDIUM
Analyzed
May 20, 2026, 08:44 PM
Security Audit — agent-trust-hub — research-integration