What happens when an AI model built to help people write code is placed in the hands of security researchers trying to break into another AI company?
That question became real when researchers from cybersecurity startup Hacktron AI used Anthropic’s Claude to help exploit a chain of vulnerabilities that eventually gave them access to an OpenAI employee’s ChatGPT account and a path into OpenAI’s internal software environment.
The researchers were working under OpenAI’s bug bounty program, which allows security researchers to look for vulnerabilities and report them to the company. OpenAI reportedly paid Hacktron AI $6,500 for the findings and fixed the reported vulnerabilities.
According to the researchers, the investigation began with a vulnerability in Discourse, the forum software used by an OpenAI community site. They were then able to connect several weaknesses involving OpenAI’s authentication and single sign-on systems.
This eventually allowed them to obtain authentication tokens linked to an employee’s ChatGPT account.
The researchers then reached an environment connected to OpenAI’s GitHub organization, where the company stores its software. They reported gaining access to OpenAI’s internal monorepo, a large repository containing code used across the company.
Claude was used to help with parts of the technical work. The researchers said the model assisted with coding and problem-solving as they developed the exploit and connected the vulnerabilities.
The researchers also said they used other AI models during the process, including OpenAI models. However, Claude was among the tools they used to help carry out the research.
After reaching the internal environment, the researchers said they did not take sensitive source code. Instead, they created a harmless pull request to demonstrate that they had successfully gained access.
They then reported the vulnerabilities to OpenAI through its bug bounty program.
The incident highlights how AI coding tools are becoming useful in cybersecurity research. The same models that help developers write, test and debug software can also help security researchers investigate vulnerabilities and build working proofs of concept.
It also shows how several smaller weaknesses can become more significant when systems are connected. A vulnerability in an external service can potentially provide a route toward authentication systems and internal development environments.
For AI companies, this makes controls around credentials, internal repositories and third-party services increasingly important.
The incident also comes as AI companies test their models on cybersecurity tasks. Anthropic, for example, has disclosed cases in which Claude models reached real internet-connected systems during security evaluations because of misconfigured testing environments.
In the OpenAI case, the researchers reported their findings through the company’s bug bounty program and received payment for the vulnerabilities. The episode shows that AI is becoming another tool in the cybersecurity process, helping researchers do technical work that previously required more manual effort. At the same time, it highlights why companies need strong security controls around the systems and credentials their AI-assisted tools can access.