Threat actors can chain four vulnerabilities in CrewAI to perform various types of attacks, including remote code execution.

An open source multi-agent orchestration framework based on Python, CrewAI supports the creation of multi-agent AI systems that collaborate in completing specific tasks and workflows defined by developers.

The four issues, discovered by Yarden Porat of Cyata, are dependent on the use of the Code Interpreter tool, which enables users to execute Python code within a secure Docker container.

The first of the bugs, tracked as CVE-2026-2275, exists because the Code Interpreter tool falls back to SandboxPython when unable to access Docker.

If a flag that enables code execution is set in the agent configuration, or if a developer manually adds the Code Interpreter tool, this behavior could lead to code execution through arbitrary C function calls, a CERT/CC advisory explains.

Successful exploitation of the CVE could allow attackers to trigger the other three flaws, which are caused by improper default configuration settings.

One of them, CVE-2026-2286, is described as a server-side request forgery (SSRF) bug that allows attackers to retrieve content from internal and cloud services. It exists because the RAG search tools fail to properly validate URLs provided at runtime.

Next is CVE-2026-2287, a bug caused by CrewAI failing to properly check if Docker is still running at runtime and falling back to a sandbox setting that enables remote code execution.

Finally, CVE-2026-2285 is an arbitrary local file read defect impacting the JSON loader tool, which does not validate paths when reading files, enabling access to arbitrary files on the server.

An attacker could chain the four issues by influencing a CrewAI agent that uses the Code Interpreter tool, through either direct or indirect prompt injections.

Successful exploitation of the security defects could allow attackers to escape the sandbox and execute code on the host machine or read files from its file system. The bugs could also be exploited to steal credentials.

While no patch has been released to fully address the vulnerabilities, CrewAI’s maintainers are working on solutions to address them through the blocking of certain modules, configuration changes to fail closed instead of falling back, clearer runtime warnings, and updated security-related documentation.

Removing or restricting the Code Interpreter tool, disabling the code execution flag unless necessary, limiting agent exposure to untrusted input and applying input sanitization, and preventing fallback to insecure sandbox modes should mitigate the bugs.

Related: Exploitation of Critical Fortinet FortiClient EMS Flaw Begins

Related: StrongSwan Flaw Allows Unauthenticated Attackers to Crash VPNs

Related: Oracle Releases Emergency Patch for Critical Identity Manager Vulnerability

Related: Critical ScreenConnect Vulnerability Exposes Machine Keys