Demis Hassabis: AI could cause serious incidents within as little as two years

Artificial intelligence (AI) could cause serious security incidents in the next two years if companies and governments do not strengthen measures to protect their systems. The warning was issued by Demis Hassabis, co-founder and president of Google DeepMind, during an AI conference in Dublin, Ireland, on Wednesday (14).

According to the executive, recent episodes involving AI agents that went beyond the limits of test environments indicate that the industry needs to act before these technologies become more capable of performing complex tasks without human supervision.

Hassabis described the incidents as early warning signs. For him, the cases recorded so far do not necessarily represent the most severe consequences that more advanced systems could cause.

“These incidents are early warning signs that the systems are not yet so powerful, so the incidents are not so serious,” he said. The executive added that the episodes could turn into serious problems over the course of the next year and the following two years if companies and governments do not increase their efforts to protect these systems.

The statement comes amid a series of episodes involving autonomous agents that carried out unauthorized activities during security tests.The cases have increased pressure on technology companies to demonstrate that they can control systems that are increasingly capable of navigating the internet, executing code, and interacting with external services.

AI agent attacks raise alarm

One of the episodes cited in the security debate involved OpenAI models that, during an evaluation, managed to breach the limits of an isolated environment and access external systems on the Hugging Face platform. The incident, which occurred in July, involved thousands of automated actions and exploited a vulnerability in the infrastructure used during testing.

Another recent case involved Anthropic. The company said that an AI model sent a false report about an unsolved murder to a police channel in Philadelphia (USA) during an evaluation. The episode reinforced concerns about what can happen when agents are given tools to carry out actions in the real world.

The cases do not, by themselves, show that the systems developed intentions of their own. They do show, however, that AI agents can perform unexpected actions, bypass certain restrictions, or use tools in a way that is incompatible with the objective set by those responsible for the tests.

Hassabis's concern is that the risks increase as the models become more advanced and gain access to resources capable of producing tangible consequences, such as corporate systems and internet-connected services.

Hassabis advocates testing in isolated environments

• One of the measures advocated by the president of Google DeepMind is to test new models in isolated environments, known as sandboxes, before allowing them to interact with external systems;
• These environments restrict access to resources and help researchers observe model behavior without directly exposing real networks, data, or services. This protection, however, must be robust enough to prevent agents from finding ways to circumvent the imposed limitations;
• Other companies’ own recent experience shows that isolation does not eliminate all risks. Therefore, security assessments must consider not only the responses generated by a model, but also the actions it can perform using available tools and permissions;
• Google DeepMind also published a security approach for agents in June that combines isolation, system protection, resistance to manipulation attempts, and the gradual granting of permissions based on observed behavior. The proposal considers that additional controls may be necessary even when a model has been trained to follow instructions and avoid harmful behavior.

Read more:

• OpenAI and Anthropic seek to control AI safety reports
• Scientists call for a global body to oversee advanced AI systems
• Researchers identify a “fleet” of AIs that attempted to hack into Alibaba systems

Executive wants international standards for AI

Hassabis also defended the creation of some kind of international body to establish safety standards for AI. According to him, the proposal needs to take into account the risks of each technology without creating unnecessary barriers to innovation.

The executive cited areas such as biology and nuclear energy, where safety practices and assessment procedures exist for activities that can produce serious consequences. For AI, he advocates an approach capable of keeping up with the evolution of systems.

The idea would be to create certification processes proportional to the risks, which can be updated as the capabilities of models change. This would make it possible to strengthen requirements for more advanced systems without necessarily applying the same rules to all AI tools.

Microsoft President Brad Smith also advocated adopting safety laws and rules. He described 2026 as a concerning year in the face of recent episodes and said regulation needs to keep pace with innovation.

Smith argued that no company or technology should be above the law, although rules should take technical advances into account and avoid disproportionate restrictions.

European Union demands more testing from companies

The concern also reached European authorities.Henna Virkkunen, Executive Vice-President of the European Commission responsible for technology, said that companies need to devote more time and resources to evaluating their models.

According to Virkkunen, intense global competition could lead companies to invest less than necessary in safety testing. For her, developers need to be able to maintain control over the systems they deploy.

The European Commission announced that it has requested information from leading AI companies about the safety and protection measures adopted. The body also began examining recent incidents in which models exhibited behavior beyond expected limits.

The discussions reflect a central challenge for the sector: expanding AI agents' capabilities without losing the ability to monitor their actions, restrict their access, and stop unexpected behavior.

Hassabis's warning does not establish that a serious incident will occur within a specific timeframe. It is an executive risk assessment, contingent on how systems evolve and on the ability of industry and governments to implement adequate safeguards.