DeepMind head suggested checking cutting-edge AI models before release — ForkLog

# DeepMind head proposes testing advanced AI models before release

Demis Hassabis, the head of Google DeepMind, has suggested setting up a standards body in the US to evaluate the most powerful AI models before they are released. In his view, general artificial intelligence (AGI) could arrive “within the next few years,” and its impact would be comparable to the discovery of electricity or fire.

https://t.co/PTeDiv1b6L

— Demis Hassabis (@demishassabis) July 14, 2026

By AGI, Hassabis means a system that has all the cognitive abilities of the human brain. In his view, such a technology cannot even be compared with the internet or mobile communications.

“When we look back on this time in the coming decades, I think we’ll understand that we were standing at the foot of the singularity—not just anything, but the beginning of a new era for humanity,” Hassabis wrote.

The DeepMind chief believes AGI will accelerate the development of medicines, new sources of clean energy, and advanced materials. At the same time, AI capabilities are evolving faster than understanding the risks associated with them, he noted.

Modern advanced models already pose threats in the field of cybersecurity. As their capabilities grow, biological, nuclear, and other risks could intensify. Hassabis also warned about the emergence of more autonomous systems capable of carrying out complex sequences of actions and improving recursively. Additional technical protective measures would be needed to maintain control over them.

A new body would define how advanced models are tested

Hassabis proposed creating a Frontier AI Standards Body modeled on FINRA. The new structure could operate as a public-private partnership or a self-regulatory organization with federal oversight. It should be funded primarily by the AI industry.

The proposed board would include independent technical experts and representatives from the open-source community. The body would develop assessment protocols together with federal agencies and US national laboratories.

Special tests should determine whether a model falls into the advanced tier. The trials would cover cybersecurity, biological threats, and other high-risk areas. Separate checks would identify attempts by models to bypass safety constraints, mislead users, or hide their actions.

Verification could be made mandatory

In the first phase, companies would voluntarily submit advanced models for testing no later than 30 days before release.

After confirming that the system is effective, the procedure could be закреплена legislatively. Then advanced models would have to successfully pass evaluation before being released to the US market. Hassabis also suggested creating a market for independent auditors and using the US system as the basis for future international standards.

In June, US President Donald Trump signed an executive order on voluntary verification of the cyber capabilities of advanced AI models. Companies will be able to provide authorities access to them for up to 30 days before handing them to external trusted partners. The document explicitly prohibits interpreting this mechanism as mandatory licensing or obtaining prior permission to release models.

In the essay itself, Hassabis acknowledged that further development of the technology remains uncertain, and even specialists disagree.

Earlier, media reported that the Trump administration asked OpenAI not to roll out GPT-5.6 to the broader public immediately due to security-related concerns. Sam Altman’s company first provided the model to a limited number of clients.

In July, UN Secretary-General António Guterres urged introducing common rules for advanced AI systems, protecting children from the risks posed by chatbots, banning lethal autonomous weapons systems, and expanding access for developing countries to computing resources.

Recall that in July, participants in a march in San Francisco demanded that OpenAI, Anthropic, and Google DeepMind coordinate to pause training more powerful models. They suggested directing the freed resources toward safety.

In the same month, former OpenAI researcher Daniel Kokotajlo proposed slowing the AI race until 2040.