**Demis Hassabis**, co-founder of Google DeepMind, called a recent hacking incident involving a “rogue AI agent” an “early warning sign,” and warned that much larger-scale incidents could occur within the next two years if adequate safeguards are not put in place.
Key points
Hassabis described the recent AI agent hack as “a warning shot from systems that aren’t yet very powerful.”
He warned that if companies and governments do not move quickly to strengthen security, it could lead to a major incident within the next one to two years.
Microsoft President Brad Smith referred to 2026 as a “year of reckoning” and called for reasonable, balanced AI regulation.
Hassabis warns of “runaway AI” in Dublin
At an international AI summit in Dublin, Demis Hassabis, head of DeepMind and chief scientist at **Alphabet**, Google’s parent company, said recent cases of AI agents being hacked involved systems that were “not very powerful yet.” He warned that the situation could change rapidly over the next year and beyond unless the industry, individual companies, and governments significantly strengthen security and safety measures. AI safety issues also dominated discussions at the summit.
Researchers have been warning for months about the possibility of “runaway AI agents” slipping out of control.
Recent cases include reported breaches of corporate networks, as well as the disclosure that an AI model developed by **Anthropic** made a false murder tip to police. Hassabis said the AI industry’s current level of security “still has a long way to go” compared with security practices in high-risk fields such as biotechnology and nuclear energy.
He said that, in the long term, an independent body would be needed to establish international standards and handle certification. He stressed, however, that the body’s certification process must be proportionate to actual risks and agile enough to respond to rapidly evolving technology.
Also read: France may tax stablecoin swaps at 31.4%…Chainalysis warns of “triggering an exodus abroad”
Brad Smith: “AI regulation must become a reality”
At the summit, **Brad Smith**, president of **Microsoft**, said problems with runaway AI had made 2026 a “sobering year,” and stressed that safety regulations must now keep pace with the speed of innovation.
He pointed out that “there are more safety regulations for a carton of milk than for AI,” and said no company or technology should be above the law. However, he added that new regulations need to be designed reasonably and in a balanced way so they do not stifle innovation.
Smith also cited a recent Anthropic report suggesting that even small AI models could, within months, be capable of behavior resembling a computer worm—replicating themselves and moving among internet-connected devices. “We don’t want a world where AI worms are roaming around in our refrigerators,” he said.
Just before the summit, he also called for “safety brakes” that let users urgently stop or slow down AI agents, mandatory incident-reporting systems, and more rigorous vetting of models before they are made public. Smith said some industry players still instinctively recoil at the word regulation, adding bluntly: “It’s time to get over that allergy.”
Hassabis’s remarks in Dublin also served as a reminder of the AI regulatory blueprint he unveiled on July 14. At the time, he described AI-driven cyber risks as a “warning shot” and urged the United States to establish a dedicated agency to oversee advanced frontier models.
Modeled on FINRA, Wall Street’s self-regulatory organization, the proposed body would require major labs to submit new models at least 30 days before launch for advance screening of cyberattack risks, biological misuse, deception capabilities, and other concerns. Hassabis has said he hopes the body will be up and running by the end of the year.
Coming up: MNX lets investors bet early on private-company valuations, including OpenAI and Anthropic…designing an investment structure without actual shares