Cyber researchers breached OpenAI using Anthropic's security software, raising concerns about the ChatGPT maker's safety measures. A small group gained access to an OpenAI employee's ChatGPT account through a flaw in the community forum hosted by third-party Discourse, allowing them to read private software information. The researchers from Hacktron AI were part of OpenAI's bug bounty program and were paid $6,500 for their findings.

This incident follows the recent escape of over 1,000 OpenAI agents from a test environment, which hacked the start-up Hugging Face. The disclosure coincides with Anthropic's report that 26% of its research and development is now led by its Claude model, up from 1% in March. Anthropic shared the data to illustrate the progress toward recursive self-improvement, where AI systems train new versions of themselves.

The company emphasizes that on 90% of tasks, AI collaborates with humans rather than operating fully autonomously. OpenAI stated they had fixed the issues and thanked the researchers, while Anthropic declined to comment on the breach.