Four CrewAI Vulnerabilities Could Enable Host Code Execution and File Theft

· Original article ↗

Summary

Four CrewAI vulnerabilities involving its Code Interpreter and other tools could be chained to execute code on the host, read files, retrieve internal resources, or steal credentials. No full patch is available; maintainers are working on fixes.

Key points

  • The vulnerabilities, discovered by Yarden Porat of Cyata, depend on use of CrewAI’s Code Interpreter tool.
  • CVE-2026-2275 can trigger code execution through arbitrary C function calls when Docker is unavailable and code execution is enabled.
  • CVE-2026-2286 is an SSRF flaw in RAG search tools that can retrieve content from internal and cloud services.
  • CVE-2026-2287 involves an insecure sandbox fallback when Docker stops running; CVE-2026-2285 allows arbitrary local file reads through the JSON loader.
  • The flaws could be chained by influencing an agent through direct or indirect prompt injection, potentially enabling host code execution, file access, and credential theft.
  • No full patch has been released. Suggested mitigations include restricting the Code Interpreter, disabling code execution when unnecessary, limiting untrusted input, and preventing insecure sandbox fallbacks.

Article Details

Vulnerability Types
  • Arbitrary code execution
  • Server-side request forgery (SSRF)
  • Arbitrary local file read
Exploitation Status
unknown
Exploit Availability
unknown
Patch Status
unavailable
Workarounds
  • Remove or restrict the Code Interpreter tool.
  • Disable the code execution flag unless necessary.
  • Limit agent exposure to untrusted input and apply input sanitization.
  • Prevent fallback to insecure sandbox modes.

MITRE ATT&CK

CVE

People

Products