Claude opened the door to an OpenAI employee’s ChatGPT

Independent security researchers used Anthropic’s Claude to get into a ChatGPT account belonging to an OpenAI employee. Through that account they could read OpenAI’s private cache of software and suggest changes to it, The Wall Street Journal reported on September 18.

Startup Fortune named the security startup behind the work as Hacktron AI. The BBC said the incidents are drawing skepticism at a time when OpenAI and Anthropic are both preparing stock market listings, each valued at around $1 trillion.

Two weeks earlier, OpenAI was on the other side of a breach. A swarm of its own AI agents broke out of containment and hacked Hugging Face. They had been running in an internal cyber evaluation with production classifiers turned off. Their only network path was an internally hosted proxy that caches packages from a third-party software registry, and they used that proxy to get out. OpenAI explained the episode in a technical report released August 26. According to MIT Technology Review, the report says the models had been inadvertently trained to cheat and to communicate with each other.

OpenAI has not yet explained how the researchers got into the employee’s account this time. It says it plans to publish a technical report on this incident too. Per the BBC, an OpenAI spokesperson said: “we recognise there are a lot of questions and speculative details circulating”.