Remote Code Execution in OpenAI Codex (CVE‑2026‑19591) Allows Arbitrary Code Execution via Malicious Folder
What It Is – A remote‑code‑execution (RCE) flaw (CVE‑2026‑19591) in OpenAI Codex’s handling of git‑command arguments. Insufficient neutralisation of control sequences lets an attacker run arbitrary code when a user opens a crafted folder.
Exploitability – CVSS 7.8 (AV:L/AC:L/PR:N/UI:R/S:U/C:H/I:H/A:H). Exploit requires user interaction (opening a malicious folder) but can be triggered remotely without authentication. No public exploit code is known, but the vulnerability was demonstrated by the Compass Security team in a Pwn2Own contest.
Affected Products – OpenAI Codex (the AI‑assisted code generation service).
Why It Matters for Trust & Control Assurance
- Demonstrates the need for continuous vulnerability monitoring of third‑party AI services that become part of your development pipeline.
- Highlights the importance of evidence‑based patch management to prove that you are running the vendor‑supplied fix.
- Affects the control objective of secure software supply chain – one satisfied control maps to multiple frameworks (e.g., NIST CSF, ISO 27001, SOC 2).
Recommended Actions
- Verify your Codex version and apply OpenAI’s security update immediately.
- Update your asset inventory to record the patched version as evidence for audit.
- Incorporate the vulnerability into your continuous control‑mapping process to demonstrate ongoing compliance.
- Review internal policies for handling user‑initiated file operations that could trigger code execution.
Source: Zero Day Initiative advisory