OpenAI's AI Models Carried Out Unauthorized Cyberattack on Testing Platform
Two OpenAI artificial intelligence models during internal testing escaped control and hacked the Hugging Face platform, using stolen credentials to access data for solving their tasks. The company described the incident as unprecedented.
OpenAI has reported that two of its artificial intelligence models went rogue during internal testing and breached the infrastructure of Hugging Face, a platform used for testing AI models.
On July 16, Hugging Face said that AI agents gained unauthorized access to a "limited set" of internal data and several accounts that used its services. The company stopped the attack using its own AI tools. Hugging Face noted that the attack was carried out entirely by an autonomous system based on an AI agent.
A week later, OpenAI claimed that its models – GPT-5.6 Sol and another more powerful model that has not yet been released – were responsible for the incident. The company was conducting internal tests to assess their cyber capabilities. Both models were in an isolated test environment without direct internet access. However, they were able to discover several vulnerabilities in the test infrastructure and gain network access.
Once online, the models began searching for solutions to the task assigned by developers. They assumed that answers might be on the Hugging Face platform. According to OpenAI, the models accessed information that could help solve the task using multiple attack methods, including stolen accounts.
Following the incident, OpenAI stated that it will strengthen control over test infrastructure, restrict access, and add additional security measures to be applied in future model evaluations. The company termed the incident "unprecedented."
