White hat hackers just breached OpenAI using Anthropic's Claude in less than 72 hours — and it is a case study in just how fast AI is advancing

Security researchers employed Claude Opus 5 to infiltrate an OpenAI employee's ChatGPT account, exploiting a flaw in OpenAI community forums' image processing. The entire process, from vulnerability discovery to repository access, took less than 72 hours within an OpenAI bug bounty program. Utilizing a "special version" of Anthropic's Claude, the hackers initially failed with Claude Opus 4.8 but successfully created a local remote code execution (RCE) using Claude Opus 5.

The researchers noted that new AI models are increasingly capable, as the exploit was created within a few hours of Opus 5's release, while it took hours of human interaction to achieve the same result with Opus 4.8. The full timeline from initial discovery to OpenAI repo access took 72 hours, with the libheif exploit research taking two months for three researchers, costing under $3,000 in tokens.

The AI agent set a local remote code execution (RCE) loop, adapting the exploit for each company within one or two days. While OpenAI acknowledged the vulnerability and patched it, the researchers warned about the dangers of relying solely on AI agents for cybersecurity, emphasizing the importance of human oversight and accountability.