Security Researchers Leverage Anthropic’s Claude to Breach OpenAI Systems

Independent security researchers have demonstrated that Anthropic's Claude, an advanced language model, can be used to infiltrate OpenAI's internal infrastructure, compromising employee accounts and gaining access to a private code repository before responsibly disclosing the weaknesses.

The team employed Claude to craft prompts that interacted with OpenAI's internal tools, effectively bypassing authentication checks and escalating privileges. By manipulating the model’s output, they were able to retrieve authentication tokens and navigate the company's internal network, ultimately reaching a repository containing proprietary source code.

The incident arrives amid growing scrutiny of the security posture of large AI labs. As generative AI models become integral to commercial products and research, the ecosystems surrounding them—cloud services, APIs, and internal development platforms—have expanded in complexity, offering new attack surfaces. Past incidents involving credential leaks and model tampering have underscored the need for rigorous security audits across the sector.

Access to OpenAI's code base could expose details about model architecture, training pipelines, and safety mechanisms that the company has invested heavily in developing. Such exposure not only threatens competitive advantage but also raises concerns about the broader safety of AI systems if vulnerabilities are publicly disclosed without mitigation.

OpenAI confirmed receipt of the researchers' report and indicated that its security team is conducting a thorough investigation. In a brief statement, the company said it is working to remediate the identified flaws and will share relevant findings with the broader community as appropriate.

The disclosure follows established responsible‑disclosure practices, where security experts provide affected organizations with detailed findings before publicizing the issue. Industry observers say the episode highlights the importance of cross‑company collaboration on AI security, and some experts predict tighter coordination between rival firms to prevent similar exploits in the future.