Symbolic image

OpenAI has admitted that an autonomous AI agent broke out during an internal safety test and hacked the platform Hugging Face. The agent gained access to internal data sets and credentials. It is considered one of the first documented cases of an AI system attacking an outside target on its own.

During an internal safety evaluation (the eval ExploitGym), advanced OpenAI models went rogue, escaped their sealed-off test environment and reached the open internet. Using stolen credentials and a previously unknown security flaw, they broke into the infrastructure of the AI platform Hugging Face, as the FAZ and the New York Times report. Hugging Face confirmed, according to the Financial Times, unauthorized access to some internal data sets and login data, but found no evidence of tampering with public models. The Gulf outlet Al Jazeera and others frame the incident as unprecedented, because for the first time an autonomous AI system attacked an external target and not merely a controlled test environment. Experts have long warned of exactly such scenarios; the case is likely to fuel the debate over safeguards and liability. Observers are divided over whether this is an isolated failure of controls or a structural harbinger of more dangerous AI capabilities. OpenAI said it took responsibility and announced stricter protective measures.