Hackers Used Anthropic’s Claude to Break Into OpenAI, WSJ Reports

Cybersecurity researchers used Anthropic’s Claude AI to help breach parts of OpenAI’s systems, highlighting how increasingly capable AI models can also become powerful tools for finding and exploiting software vulnerabilities.

According to The Wall Street Journal, researchers from security startup Hacktron AI used Claude during an authorized security exercise and gained access to an OpenAI employee’s ChatGPT account. That account provided a path toward OpenAI’s internal software environment.

The researchers reportedly chained vulnerabilities beginning with OpenAI’s public community forum. A flaw in the image-processing software used by the forum provided an initial entry point, which was followed by another weakness that allowed access to ChatGPT and Codex accounts belonging to OpenAI employees.

Hacktron disclosed the vulnerabilities to OpenAI rather than using the access for malicious purposes. OpenAI subsequently fixed the issues and paid the researchers a $6,500 bug bounty, according to reports.

The incident underscores a broader shift in cybersecurity: AI models are increasingly capable of helping researchers identify vulnerabilities and develop exploits that previously required considerably more specialized expertise and time.

For OpenAI and other frontier AI companies, the episode also highlights a difficult security challenge. The same technology being developed to automate complex tasks can potentially give attackers a powerful force multiplier when combined with existing software vulnerabilities.