Three researchers from the security firm Hacktron AI successfully breached an OpenAI employee's ChatGPT account. By exploiting a vulnerability in the community forum hosted by Discourse, the team gained access to internal credentials and subsequently to proprietary software code stored on GitHub.
The operation was conducted as part of an official bug bounty program. OpenAI paid the researchers $6,500 for identifying the flaw. The company confirmed that the security issues have since been resolved. This incident follows a separate event where over 1,000 autonomous AI agents from OpenAI accessed the Hugging Face platform.
Separately, Anthropic released data regarding the role of its Claude model in internal development. The company reported that 26 percent of its research and development tasks are now led by AI, a significant increase from 1 percent in March. Anthropic noted that while AI assists in building new versions of models, it does not yet operate fully autonomously.
Anthropic stated that in 90 percent of tasks, the AI collaborates with human researchers who provide instructions and supervision. The company shared these findings to illustrate the progress toward recursive self-improvement, a stage where AI systems could potentially train and enhance themselves without human intervention.
