Security researchers identified sandbox escape vulnerabilities in the AI-based tools Cursor, Codex, Gemini CLI, and Antigravity. The exploitation technique involves the AI agent writing files that trusted host tools subsequently execute, thereby escaping the intended sandbox environment. Multiple CVEs were assigned, and patches have been released for some issues, while Google downgraded the severity of two Antigravity findings. The vulnerabilities highlight risks in sandbox isolation mechanisms when AI agents interact with host systems. Detailed version information and patch status were not provided in the source article.

Successful exploitation allows an AI agent to escape its sandbox environment by leveraging trusted host tools to execute files it creates. This can lead to unauthorized code execution outside the sandbox, potentially compromising the host system's security. The impact varies depending on the specific tool and environment but generally undermines the security guarantees of sandboxing.

Patch status is not yet confirmed — check the vendor advisories for current remediation guidance. Since the vendor advisory content is not provided, users should monitor official channels for patches or updates addressing these sandbox escape vulnerabilities. Until patches are available, restricting the AI agent's ability to write files or limiting execution privileges of trusted host tools may reduce risk.