Researchers use Anthropic Claude to breach OpenAI employee accounts

U.S. cybersecurity researchers from a startup have demonstrated a breach of OpenAI’s internal systems by exploiting employee ChatGPT accounts, according to a report released by the team. Using Anthropic’s Claude chatbot as part of their methodology, the researchers were able to gain access to the software cache of the compromised accounts, a step that could potentially allow further infiltration of OpenAI’s infrastructure. The investigators described the theoretical reach of the attack as “huge,” indicating that the compromised credentials might have opened pathways to additional data and services beyond the initial cache.

OpenAI, which has faced heightened scrutiny over its security posture, confirmed that the incident is under investigation but declined to comment on specific details. The breach underscores ongoing challenges for AI firms in safeguarding access to powerful language models and associated tools, especially as third‑party services like Claude become integrated into research workflows. The company has pledged to strengthen its authentication mechanisms and monitor for any further unauthorized activity.