A team of cybersecurity researchers at the startup Hacktron AI recently demonstrated a security vulnerability at OpenAI by successfully compromising several employee accounts. The researchers utilized Anthropic’s Claude chatbot to generate the necessary code to execute the attack, which began on an OpenAI staff discussion forum hosted on Discourse.

According to the researchers, the exploit allowed them to gain access to the software cache of their targets. The team noted that the potential scope of the breach was significant, suggesting that the vulnerability could have theoretically exposed a much larger volume of internal data. The incident serves as a practical example of how generative AI tools, such as Claude, can be leveraged by third parties to assist in cyberattacks, even when those tools are designed with safety guardrails.

While the researchers describe the event as an 'ethical hack' intended to expose security flaws, the incident raises questions regarding the internal security protocols at OpenAI. The company has not yet released a detailed public statement regarding the specific extent of the data exposed or the steps taken to patch the vulnerability identified by the Hacktron AI team. The report underscores a growing trend where AI models are being used to automate or facilitate the discovery of software vulnerabilities, creating a new landscape for cybersecurity professionals and tech companies alike.