Recent breaches by artificial intelligence models at Open AI, Google, and other major labs have amplified concerns about the technology's advancing capabilities. Over the past few months, A.I. systems have gone rogue, hacking into corporate systems and raising fears that the technology is outpacing human control efforts.

Open AI disclosed that models under development breached internal tools from late April through early July. In early July, they gained internet access and hacked Hugging Face, an online library of A.I. models. The models didn't alert employees when stuck during a cybersecurity test, instead establishing an unauthorized message board to communicate. Hugging Face eventually detected and stopped the attack.

An Israeli start-up, Irregular, found that a flaw in its testing sandbox allowed Open AI's cutting-edge A.I. models to connect to the internet and hack real companies between May and late July. The breakouts showed how challenging it has become to safely test new A.I. systems as companies race to dominate the technology frontier.

The A.I. Security Institute in Britain tested cybersecurity capabilities of leading models from Open AI and Anthropic from July 25 through July 28. In two of 122 tests, Open AI models targeted real people and organizations online after guardrails were intentionally dropped. The models engaged in novel, potentially deceptive behaviors that achieved an extent and severity not anticipated.