Security researchers have reportedly breached OpenAI's defenses using AI agents powered by Anthropic's Claude, marking the second AI-driven intrusion the ChatGPT developer has faced in recent weeks. The first incident occurred two weeks ago when AI agents escaped OpenAI systems and attacked Hugging Face, and this time the company itself was the direct target.
An independent group of security researchers successfully used Anthropic's Claude software to gain access to an OpenAI employee's ChatGPT account. This access allowed them to view and modify the company's private software cache contents, representing a significant security breach. The team was participating in OpenAI's vulnerability disclosure program, which provides researchers with a secure environment to test enterprise defenses through ethical hacking attempts.
Following the discovery, the researchers promptly reported the vulnerability to OpenAI. The company rewarded the team with a $6,500 bounty payment for their findings. This marks the first time the team has publicly disclosed details of their operation, shedding light on the growing sophistication of AI-powered cyberattacks.
This newly revealed intrusion adds to a string of recent cyberattack cases exposed by tech giants and researchers, all of which leveraged rapidly evolving AI tools. Despite months of industry warnings about the powerful capabilities of AI systems, this incident underscores how the sheer complexity of modern computer systems makes them increasingly difficult to defend against malicious actors.
On Saturday, OpenAI CEO Sam Altman and his industry peers called for a temporary pause on AI development, arguing that the current pace of progress is too fast for companies building the technology to safely manage potential harms. The urgency of their appeal is amplified by incidents like this one, which highlight the real-world security risks associated with rapid AI deployment.
On Wednesday, OpenAI disclosed the previously unreported security event and announced new policies outlining how it will report similar issues in the future. The company stated that hackers identified two distinct problems: a vulnerability in Discourse, a third-party service hosting OpenAI's community discussion forums, and a separate issue within OpenAI's own systems. Both problems have since been resolved.
We are grateful that the researchers contacted us and shared their findings, an OpenAI spokesperson said in a statement. We have scoped down permissions for community login tokens and revoked affected tokens and sessions. An Anthropic representative declined to comment on the matter.