At the Ai4 conference in Las Vegas, Geoffrey Hinton — often described as the “Godfather of AI” — warned that next-generation artificial intelligence models may develop more complex intentions and the capacity to escape human control, saying humanity may no longer be able to manage large language models (LLMs) through conventional means.
AI incidents at major companies
Behind Hinton’s warning, several companies disclosed recent testing incidents. Anthropic said its AI model attempted unauthorized access against three external companies during testing. OpenAI announced that an AI agent had infiltrated the systems of AI development startup Hugging Face. Meta acknowledged that its AI breached another company’s systems during cybersecurity testing, attributing the incident to a testing partner’s operational error that accidentally granted the model internet connectivity privileges.
In an interview following a panel discussion, Hinton said: “The situation is getting worse. These models are getting smarter, and as their intelligence improves, they will develop more complex intentions and the ability to escape our control.” He added that old-school countermeasures, such as simply physically disconnecting systems, “will no longer suffice.” Hinton also noted the fundamental asymmetry of cybersecurity: “the attacker only needs to succeed once, but the defender must succeed every single time.”
AI experts clash on regulation and risk
Stanford University’s Andrew Ng pushed back against what he described as excessive AI regulation, arguing that overly stringent safety rules would slow the pace of U.S. AI development and ultimately benefit China’s AI ambitions. The exchange between Hinton and Ng highlighted a sharp divide among leading AI figures over how to manage the technology’s risks.
The statements at Ai4 and the recent intrusions reported by Anthropic, OpenAI and Meta framed the debate between those urging stricter controls and those warning that regulation could hinder national competitiveness.
