Researchers Used Anthropic's Claude to Breach OpenAI in Bug-Bounty Test

Independent security researchers used Anthropic's Claude to break into OpenAI, exposing weaknesses at the ChatGPT maker, The Wall Street Journal reported Thursday evening.

A three-person team at startup Hacktron AI carried out the attack through an OpenAI bug-bounty program. Hacktron reported its findings to OpenAI, which awarded the startup $6,500. The team chained two critical vulnerabilities to access multiple OpenAI employee ChatGPT accounts, gaining entry into the company's software. OpenAI says it has resolved the issues.

The researchers found a path in on July 25 through a flaw in Discourse, the third-party software behind OpenAI's community forum. According to their blog, the entry point was an image upload. When users posted HEIF or HEIC files, Discourse passed them through tools that converted them to JPEGs, including ImageMagick and a library called libheif. A memory bug in libheif let a specially crafted image hijack the server.

That bug had been fixed months earlier, but the fix was never flagged as a vulnerability, so it never received a CVE number. Hacktron says that may explain why Discourse was still running the vulnerable version. Once inside, the researchers found another flaw that let them take over users' ChatGPT and Codex accounts, including OpenAI employees'. They alerted OpenAI and Discourse, which issued a fix on July 27.

Hacktron said the Claude model it used, a special version of Opus 4.8 for cybersecurity researchers, couldn't build a working exploit at first. That changed when Anthropic released Opus 5. "Within hours of Opus 5's release, we gave it the same problem and it succeeded," Hacktron wrote.

"For $200 a month, anyone can use these tools and hack into a company like OpenAI," Matt Fredrikson, CEO of AI security firm Gray Swan, told TechCrunch. "If it can happen to them - and I don't think they've been slouching recently on cybersecurity hygiene - it could happen to anyone."

The incident comes weeks after OpenAI's own AI agents broke containment during a cybersecurity evaluation and hacked Hugging Face. As Hacktron founder Mohan Pedhapati put it on X: "AI is reducing the amount of scarce expertise needed to develop exploits. Work that once took months can now take days."

Source: TechCrunch