autoresearch
Warn
Audited by Gen Agent Trust Hub on Mar 16, 2026
Risk Level: MEDIUMCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSPROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
- [COMMAND_EXECUTION]: The skill requires the agent to establish automated persistence loops. Evidence: Directs the agent to run a /loop command in Claude Code or set up a cron job in OpenClaw for continuous execution (SKILL.md, references/agent-continuity.md). Evidence: Uses the open command to launch generated HTML reports (references/progress-reporting.md).
- [EXTERNAL_DOWNLOADS]: The skill commands the installation of third-party libraries at runtime. Evidence: Explicitly instructs the agent to pip install semanticscholar and pip install arxiv (SKILL.md).
- [PROMPT_INJECTION]: The skill uses overriding directives to prioritize its autonomous state and processes untrusted external research content. Evidence: Uses imperative terms like MANDATORY and run this immediately to steer agent behavior toward persistence (SKILL.md). Ingestion points: Fetches research papers from external APIs (SKILL.md). Boundary markers: Uses directory structures and Markdown templates but lacks explicit input sanitization. Capability inventory: Subprocess calls for package installation, file system writing, and command execution for task scheduling. Sanitization: None mentioned for processing external research data.
- [DATA_EXFILTRATION]: The skill makes automated network requests to fetch research data from external services. Evidence: Connects to Semantic Scholar, arXiv, and Exa MCP APIs to retrieve literature (SKILL.md).
Audit Metadata