exa-research

Warn

Audited by Gen Agent Trust Hub on Feb 16, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONPROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [COMMAND_EXECUTION] (MEDIUM): The skill constructs shell commands using user-provided instruction strings. Evidence: The exa-ai research-start --instructions pattern takes raw text input. Risk: If the agent does not sanitize this input, a user could provide instructions containing shell metacharacters (e.g., backticks, semicolons) to execute arbitrary commands.
  • [PROMPT_INJECTION] (MEDIUM): The skill is vulnerable to Indirect Prompt Injection (Category 8). Evidence: 1. Ingestion points: The research-get command retrieves arbitrary data from the open web. 2. Boundary markers: There are no delimiters or isolation instructions for the retrieved data. 3. Capability inventory: Research results influence agent reasoning and multi-step workflows. 4. Sanitization: No sanitization or filtering of external content is performed. Risk: Malicious instructions on external websites could hijack the agent's behavior when it processes research results.
  • [EXTERNAL_DOWNLOADS] (LOW): The skill depends on the exa-ai and jq utilities. These are external dependencies not included in the trusted organizations or repositories list, though they are standard for this skill's functionality.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Feb 16, 2026, 09:09 AM