Researchers used Claude to hack OpenAI, reaching GitHub code

The OpenAI disclosure, first reported by The Wall Street Journal, centered on how a third-party forum setup could lead to an employee's ChatGPT account and internal GitHub code. OpenAI thanked the researchers for sharing their findings and said it fixed the issues, while Anthropic declined to comment and Hacktron did not immediately respond.

The same Thursday, Anthropic published data showing its own reliance on AI in model development. It said 26 percent of research and development work was "led by" Claude, up from 1 percent in March, meaning AI completed the majority of tasks under human instruction and supervision. Anthropic said it shared the data to help the public understand how close the world is to recursive self-improvement, the point where AI can train and improve itself or new models, a threshold tied to concerns about losing human control. It added that its models did not yet operate fully autonomously for any of the research studied, and that on 90 percent of tasks AI collaborates with a human and does large chunks of work.

Taken together, the two threads point to a security question that may prove harder to answer than the forum fix itself: as AI systems take on more of the work of building the next version of themselves, the accounts and code they touch appear to sit behind third-party services that can be misconfigured. For OpenAI's security staff, the test is likely whether such third-party links to internal code are removed or merely patched; for Anthropic, the figure to watch is how the share of AI-led R&D work moves beyond the March 1 percent baseline.