Researchers from Hacktron successfully breached OpenAI systems in July 2026 as part of the company's bug bounty program. The entire process, from initial vulnerability discovery to resolution, took less than 72 hours.
The attack chain began with a heap overflow vulnerability in the libheif library on the Discourse forum platform. This allowed for remote code execution, which the researchers then combined with a single sign-on misconfiguration to impersonate an OpenAI employee.
By hijacking session tokens, the team gained access to internal corporate systems, including GitHub, Slack, and email. To demonstrate the breach, they initiated an internal pull request within the company's private codebase.
The researchers utilized Anthropic's Claude Opus 5 model to analyze server data and generate the exploit code for the heap buffer overflow. They used a cybersecurity-configured version of the model that permits specific research activities.
Following the disclosure, both OpenAI and Discourse patched the identified vulnerabilities. OpenAI awarded the researchers a bounty of $6 500 for their findings, noting that no source code was downloaded during the exercise.
