Claude Opus 5: How AI Was Used to Hack OpenAI

Security researchers from Hacktron AI successfully used Anthropic’s Claude to breach OpenAI's infrastructure. According to Ars Technica and TechCrunch, the attack leveraged a $6,500 bug bounty program and targeted vulnerabilities in OpenAI's community forum software.

How did Claude execute the security breach?

The attack began with a mundane upload of an HEIF image file to OpenAI's community forum powered by Discourse. TechCrunch reports that the file passed through ImageMagick and a library called libheif, exposing an unpatched memory bug. Although the bug was previously fixed by developers, it lacked a formal CVE tracking number, leaving Discourse running a vulnerable version.

What was the difference between Claude versions?

According to TechCrunch, the security team initially used a special cybersecurity version of Opus 4.8, which struggled across multiple sessions to build a working exploit. However, within hours of Anthropic releasing Opus 5, the team fed it the exact same problem, and it succeeded.

What data was accessed during the incident?

Ars Technica notes that the exploit allowed researchers to gain access to internal sign-ons and an OpenAI employee's ChatGPT account. Because this account was connected to OpenAI's internal code through GitHub, the team gained entry to sensitive software information before reporting the flaws to OpenAI, which quickly issued fixes.