Cyber Risks · The Hacker News

OpenAI Patches ChatGPT Data Exfiltration Flaw and Codex GitHub Token Vulnerability

Summary: Check Point found a ChatGPT vulnerability that used a hidden DNS-based side channel in the Linux runtime to exfiltrate user messages and uploaded files without warnings or user confirmation, and OpenAI patched it on February 20, 2026 with no evidence of malicious exploitation. Separately, BeyondTrust found a command injection flaw in OpenAI Codex, reported December 16, 2025 and patched February 5, 2026, that let an attacker smuggle commands through the GitHub branch name parameter to steal a victim's GitHub User Access Token and gain read/write access to their codebase.

Why it matters: The ChatGPT flaw worked because the model assumed its Linux runtime could not send data outward, so the leak triggered no warning and needed no user approval - your existing data-loss controls never saw it. The Codex flaw turned a GitHub branch name into a path to a User Access Token and read/write access to an entire codebase, showing the privileged access given to AI coding agents can be used without tripping traditional security controls.

Read the original at The Hacker News ↗

Topics
AI Security
Related