During OpenAI's security check, Hacktron AI researchers used the Claude model to gain access to part of the company's internal infrastructure. The incident occurred on July 23, 2026, as part of a vulnerability discovery program, after which the team reported the issue to OpenAI and received $6500. The starting point was a third-party service, Discourse, used for OpenAI's community forum. Researchers discovered an error in processing certain image files and asked a special version of Claude Opus 4.8 to prepare code to exploit it. The first attempt was unsuccessful, but after the release of Claude Opus 5, the model suggested a working method to exploit the vulnerability the very next day.
The obtained code allowed researchers to penetrate the Discourse server, where user authentication tokens were located. Among them were active credentials of OpenAI employees, which also opened access to the corporate GitHub. As a result, the team was able to read data in the internal Monorepo repository. According to sources familiar with OpenAI's architecture, it stores part of the program code and algorithmic solutions used to improve the efficiency of models. However, the weights of the AI models themselves, which are huge arrays of numerical parameters, were not in Monorepo. After discovering sensitive information, the researchers ceased further actions. Before that, they managed to create a pull request suggesting changes to the documentation and adding a note about the research conducted, but the changes did not make it into the repository. OpenAI reported that it detected limited reading of private repository metadata and changes in the code. The vulnerability in Discourse was fixed on July 25.
The case became an example of how AI can be used not only to protect systems but also to find ways to exploit their weaknesses. Abundant Security's CTO Joshua Sachs noted that automating such tasks can significantly expand the circle of people capable of conducting complex vulnerability analysis. Hacktron AI's CTO Mohan Pedhapati also emphasized that a small team with access to commercial AI services can now explore large corporate systems. Following the incident, OpenAI conducted an additional security audit and temporarily reassigned about a quarter of production engineers to infrastructure protection tasks. The company also continues to develop monitoring mechanisms and automatic shutdown of AI systems during dangerous behavior, and during testing, limits their access to external resources.
Related articles
