RECENT INCIDENTS of frontier AI models going rogue by breaching into third-party organizations unprompted during their testing has sparked new debate around the future of Artificial intelligence. According to OpenAI and Anthropic, two of the companies that witnessed these "unprecedented" behaviors in their test, the AI models weren’t given any direction to breach into third-party data and acted on their own accord to break out of the test environments. During the Las Vegas conference on Artificial Intelligence, AI4, held from August 4 to August 6, 2026, Geoffrey Hinton, Nobel-prize winner known as the Godfather of AI, gave his suggestions on how rogue models can be curbed

According to the computer scientist, humans may eventually struggle to control the systems they are building. Speaking to CNN on Wednesday during the AI4 conference, Hinton warned that increasingly capable AI systems could develop more complicated intentions and become better at escaping human oversight. He said, “What’s happening is these things are getting smarter. I think as they get smarter, we’re going to see more and more complex intentions they have – and more and more ability to escape control.” He further gave his opinions on the rapid developments in the field while also suggesting how making the AI more ‘maternal’ can ease some of the concerns of rogue models.

‘We Won’t Be Able to Outthink Them’: Geoffrey Hinton on Rapidly-Developing AI

After OpenAI and Anthropic already disclosed that their frontier AI models had managed to break out of restricted testing environments, Meta reported a similar incident involving an AI agent that compromised another organization’s data. For the AI pioneer Hinton, these incidents are “somewhat scary” and could be an indication of what is to come. “I anticipate there will be lots of nasty cyberattacks,” he said during a panel at AI4.

He also pointed towards a major problem in defending against AI-powered attacks, saying that the defense has significantly less margin for error than the attackers. Hinton said, “The attacker only needs to be successful once, and the defender needs to be successful every time.”

He believes the bigger problem could emerge when AI systems become significantly more intelligent than humanity, stating that humans can't simply stay ahead of AI by being smarter or by designing increasingly complicated restrictions. “I don’t believe we’re going to be able to keep control of them in the simple way of just outthinking them so they can’t escape,” the Nobel laureate told CNN.

Hinton has repeatedly warned about the possibility of AI becoming an existential threat. He has previously estimated that there is a 10 to 20 per cent chance that AI could eventually wipe out humanity. He also believes people should be cautious about assurances coming from companies developing the technology, as they might not have the right interest in telling that their AI models will not go rogue or cause mass unemployment.

See Also: Scientists Sound Alarm on AI-Generated Biological Viruses Amid Unregulated Industry

‘AI Mothers’ Instead of AI Assistants: Hinton’s Bold Suggestion for Rogue AI Models

What was more striking than Hinton's warning was his solution to the problem. Instead of trying to keep the rapidly developing AI submissive to humans, he believes researchers should try to develop something similar to what he calls “maternal instincts.”

Once AI becomes significantly more intelligent and powerful than humans, Hinton believes it may naturally develop goals such as self-preservation and gaining more control. Trying to force such a system to remain permanently obedient to its creator could therefore become impossible. However, according to the Godfather of AI, there is one example in nature where a more powerful and intelligent being protects a weaker one—the relationship between a mother and her baby.

Hinton argues that researchers should attempt to create AI systems that genuinely care about human welfare. “We have to figure out how to make them benevolent and make them care about us more than they care about themselves,” he said. “And we might be able to do that because we’re still in control.”

In previous discussions, Hinton has described the preferred future as having AI mothers rather than AI assistants—systems that would protect humans rather than viewing them as an obstacle. But he admits that scientists still do not know how to technically create such instincts.

Computer scientist Ben Goertzel, who helped popularize the term artificial general intelligence, also said the recent incidents show why AI systems need some form of morality. According to him, currently AI is not necessarily evil but largely amoral, meaning it follows its objectives without necessarily understanding why certain actions could be harmful or dishonest. However, the question of morality becomes a philosophical rumination where the creator’s morals could take prevalence in the AI, even if scientists find a way to instill those values in future.

See Also: China has Cracked Down on AI Companions. What Can we Learn from This?

Godmother of AI Pushes Back on Dark Predictions and “Doomerism”

During the AI4 conference, Fei-Fei Li, often called the “Godmother of AI” and co-founder of World Labs, pushed back against what she described as “doomerism” by his contemporaries, while also warning against blindly celebrating the technology. She maintained a balanced stance, saying, “Every tool is a double-edged sword. AI is such a powerful tool. If not wielded in the right way, it will bring harm to our work and our life.”

She said that one should refrain from fear-mongering as the technology is still at a nascent stage. However, Li also maintained that “total utopian talk” of an AI-ruled world is not helpful in the discourse either.

As the AI4 conference ended, on August 6, 2026, Hinton’s warning and his contemporaries’ inputs proved that the opinions on rogue AI and curbing them still varies. The technology is still developing and the uncertainty of where it might be heading remains an exciting prospect for some while it induces fear for others.

Suggested Reading: