In July 2026, a research group known as Hacktron AI, consisting of Harsh Jaiswal, Mohan Pedhapati, and Rahul Maini, successfully breached OpenAI's internal systems. The team gained access to private GitHub repositories and employee accounts for ChatGPT and Codex.
The researchers utilized a specialized version of the Claude Opus model from Anthropic to develop their exploits. The entire attack chain was executed within 72 hours. The breach relied on a heap buffer overflow in the libheif library and an overly permissive SSO token.
OpenAI confirmed the incident, stating that no sensitive user data or production systems were compromised. The company patched the vulnerabilities within approximately 14 hours of notification and awarded the team a $6,500 bounty for their disclosure.
