Critical Vulnerabilities Found in OpenAI Codex Sandbox

• •Heapjack allows remote code execution via a shared memory heap in Codex Desktop.
• •Overpatch enables unauthorized disk access through the Codex CLI patch tool.
• •Both vulnerabilities were reported on August 12, 2026, and fixed within eight days.

Security researchers discovered two significant vulnerabilities in the OpenAI Codex sandbox, reported on August 12, 2026, and patched within eight days. The first vulnerability, named Heapjack, allows remote code execution by exploiting a shared memory heap in the Codex Desktop application, enabling untrusted code to access trusted context tokens. The second vulnerability, Overpatch, affects the open-source Codex CLI, allowing unauthorized write access to the entire disk through a patch tool. These vulnerabilities could potentially compromise developer machines and systems running Codex. Both flaws were addressed by OpenAI shortly after being reported. The Codex tool is widely used for coding assistance, making these vulnerabilities particularly concerning for developers. The researchers emphasized the lack of opt-in mechanisms for security features in Codex. The vulnerabilities highlight the importance of rigorous security practices in AI development tools.