How Hacktron AI Breached OpenAI Using Claude Opus 5 in Under 72 Hours

A team of three cybersecurity researchers from startup Hacktron AI recently demonstrated how advanced AI models can accelerate vulnerability exploitation. In an authorized security assessment, the team used Anthropic's Claude Opus 5 to chain two vulnerabilities and gain a path into OpenAI's private GitHub environment in under 72 hours.

Here is a point-wise summary of the operation and its significance for enterprise cybersecurity:

• The operation began on OpenAI's public discussion forum, which operates on the widely used open-source Discourse platform.

• The researchers found that uploading images in Apple's default HEIC and HEIF formats triggered a heap buffer overflow vulnerability in libheif, a native decoding library used by the forum's image conversion pipeline.

• The team initially tried using Claude Opus 4.8 to write an exploit, but the model struggled to bypass standard memory protections like address space layout randomization.

• When Anthropic released Claude Opus 5, the researchers fed it the same problem, and the newer model produced a working exploit within just a few hours.

• Uploading a specially crafted image allowed the researchers to execute remote code on the forum's server, granting them their initial foothold.

• The team then discovered a critical single sign-on misconfiguration: session tokens issued for the community help forum remained valid for unrelated production applications, including ChatGPT and Codex.

• Exploiting this token reuse flaw allowed the researchers to move laterally, compromise an OpenAI employee's Codex account, and reach the company's internal code repository.

• The researchers halted their test after opening a harmless pull request in the internal repository, proving they had access without actually viewing sensitive source code.

• OpenAI resolved both the image processing vulnerability and the session token configuration issue, awarding the Hacktron team a $6,500 bug bounty for responsible disclosure.

• This exercise highlights how advanced AI tools can significantly compress the timeframe required to develop complex exploits, lowering the barrier to entry for sophisticated cyberattacks.

• Security experts emphasize that while AI accelerated the exploit development, the breach ultimately relied on standard misconfigurations in third-party dependencies and authentication flows, underscoring the need for organizations to strictly audit session token lifetimes.