Breaking

OpenAI Codex CLI Sandbox Escape Lets Researchers Execute Arbitrary Commands on Host Systems

September 20, 2026 at 08:00 EDT

Security
Software Dev & Coding
AI Agents

Security researchers have discovered multiple sandbox escape vulnerabilities in OpenAI's Codex CLI and three other leading AI coding agents, demonstrating how indirect attacks through trusted development workflows can bypass security boundaries designed to isolate AI-generated code.

AI Security Alert

Codex Sandbox Escape Chain

Researchers discovered multiple attack vectors allowing host command execution despite sandbox isolation in OpenAI's Codex CLI and three other AI coding agents.

Attack Vectors Discovered

Command Allowlist Flaw

Git show trusted by name; parameters could modify files

Config Directory Bypass

Apply_patch creates .codex folder without approval

Notify Command Escape

Executes outside sandbox despite isolation enabled

Docker Socket Abuse

Communicates with privileged host service

Scope of Vulnerability

Vulnerabilities Found

AI Tools Affected

Vendors Implicated

Patch Status

OpenAI Codex CLI: Update to version 0.95.0 or later. All tools: Most vendors have released patches; check your version and enable updates immediately.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.