AI-Powered Exploit Breaches OpenAI's Code Repository in Under 72 Hours

A team of researchers from Hacktron AI successfully breached OpenAI's private code repository using Anthropic's Claude. The exploit, which took under 72 hours to develop, was made possible by switching to Claude Opus 5, released just days prior. This vulnerability allowed the researchers to access sensitive source code and even opened a harmless pull request to prove their presence.

The team first tried using Claude Opus 4.8 but found it unreliable for the task. They then switched to Claude Opus 5, which produced a working ARM64 exploit within hours. The researchers adapted this exploit for x86-64 and jemalloc environments, demonstrating the potential for rapid development of sophisticated exploits.

This incident highlights the rapidly evolving landscape of AI-powered vulnerability exploitation. The researchers' success in breaching OpenAI's system has significant implications for the crypto space, where similar techniques are being used to post malware instructions on public blockchains at an alarming rate. Chainalysis reported a 440% increase in malicious onchain writes compared to last year, with state-linked operators from North Korea and Iran accounting for two-thirds of new activity.

The incident also raises concerns about the increasing reliance on AI-powered tools for vulnerability exploitation. As more sophisticated models become available, the potential for rapid development of complex exploits grows exponentially. This trend is likely to continue, with Defillama recording April 2026 as crypto's most-hacked month on record.