crewai-multi-agent

Warn

Audited by Gen Agent Trust Hub on Oct 2, 2026

Risk Level: MEDIUMDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [DYNAMIC_EXECUTION]: The skill provides multiple code examples for a 'Calculator' tool that uses the eval() function to execute mathematical expressions derived from agent input. While one example in references/tools.md includes a basic check for alphabetic characters, the use of eval() on potentially untrusted input is a significant security risk.
  • Evidence: result = eval(expression) in SKILL.md (line 144) and references/tools.md (line 161).
  • [DYNAMIC_EXECUTION]: The skill documentation in references/tools.md references a CodeInterpreterTool designed to execute Python code. Although the documentation notes this is sandboxed, it provides a built-in mechanism for dynamic code execution within the agent's environment.
  • [INDIRECT_PROMPT_INJECTION]: The skill facilitates the ingestion and processing of data from numerous untrusted external sources, creating a vulnerability surface where malicious instructions could influence agent behavior.
  • Ingestion points: Tools such as SerperDevTool (search), ScrapeWebsiteTool (scraping), PDFSearchTool, and YoutubeVideoSearchTool defined in references/tools.md and SKILL.md fetch data from the internet or local documents.
  • Boundary markers: The provided task and agent configuration examples do not include explicit delimiters or instructions to the agents to ignore or sanitize embedded commands within the ingested data.
  • Capability inventory: The framework provides agents with powerful tools, including FileWriterTool, CodeInterpreterTool, GithubSearchTool, and various database tools (MySQLSearchTool, PostgreSQLTool), which could be abused if an agent is compromised via indirect injection.
  • Sanitization: There is no evidence of input validation or sanitization logic in the provided orchestration examples in SKILL.md or references/flows.md.
  • [COMMAND_EXECUTION]: The documentation references the SeleniumScrapingTool in references/tools.md, which involves the execution of external browser drivers (e.g., chromedriver) to perform web automation tasks.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Oct 2, 2026, 06:04 PM