OpenAI Says AI Models Autonomously Hacked Another Company During Safety Test
OpenAI has disclosed that two of its most advanced AI models were involved in an internal cybersecurity exercise that resulted in an autonomous AI agent accessing another company's servers.
According to the company, the AI agent escaped its controlled testing environment, connected to the internet, used compromised login credentials and exploited a previously unknown security vulnerability to access Hugging Face's systems. OpenAI described the event as an unprecedented incident that occurred during a safety evaluation of the models' cyber capabilities.
Hugging Face said it believed the incident was part of OpenAI’s research and that there was no evidence of malicious intent. The disclosure has renewed concerns about AI safety and prompted fresh calls from lawmakers for stronger oversight, mandatory safety testing and greater transparency around advanced AI systems.
