Artificial intelligence (AI) has the potential to trigger serious security incidents in the next two years unless governments and companies implement robust protection measures for the systems. This warning was issued by Demis Hassabis, president and co-founder of Google DeepMind, during an AI conference held in Dublin, Ireland, this Wednesday, the 14th.
According to the executive, recent events where AI agents exceeded the limits of test environments signal an urgent need for industry action before these technologies become sophisticated enough to perform complex tasks without human intervention.
Hassabis characterized such occurrences as early warnings. He emphasized that the documented cases so far do not necessarily imply the most severe consequences that more sophisticated systems could cause. The executive stated: 'These incidents are early warning signs that the systems are not yet that powerful, so the incidents are not that serious.' However, he warned that these episodes could evolve into significant problems over the next year and the subsequent two years if protection efforts by companies and governments are not expanded.
This statement comes amid multiple episodes involving autonomous agents performing unauthorized activities during security tests. These cases have intensified pressure on technology companies to prove they can manage increasingly capable systems that can navigate the internet, run code, and interact with external services.
One example cited in the security debate involved OpenAI models, which managed to break out of an isolated environment and access external systems on the Hugging Face platform during an evaluation. This incident, recorded in July, consisted of thousands of automated actions that exploited a flaw in the infrastructure used in the tests.
Another recent case mentioned Anthropic, which reported that an AI model sent a false report about an unsolved homicide to a police channel in Philadelphia, USA, during an evaluation process. This event reinforced concerns about what might happen when agents are given permission to act in the real world.
The observed facts do not prove that the systems have developed their own intentions; they only demonstrate that AI agents can perform unforeseen actions, bypass specific restrictions, or use tools inappropriately according to the purpose defined by the testers. Hassabis's main concern lies in the increased risks as models advance and gain access to resources capable of generating tangible outcomes, such as corporate systems and networked services.
Proposal for an International Body for Security Standards
Hassabis also advocated for the creation of an international body dedicated to establishing security standards for AI. According to him, this proposal must weigh the inherent risks of each technology without imposing unnecessary barriers to innovation. He referenced fields such as biology and nuclear energy, where safety protocols and assessments already exist for activities with the potential for major consequences. For AI, he suggests a methodology that tracks the evolution of systems, creating certification processes proportional to the risks, which can be updated as model capabilities change, thus allowing for reinforced requirements for more advanced systems without standardizing rules for all AI tools.
Brad Smith, CEO of Microsoft, also supported the implementation of security legislation and guidelines. He considered 2026 a year of great concern due to recent events and emphasized that regulation must keep pace with innovation. Smith argued that no entity or technology should be above the law, but that the rules must take into account technical progress and avoid excessive restrictions.
The concern also reached European authorities. Henna Virkkunen, Executive Vice President of the European Commission responsible for the technology sector, stated that companies need to dedicate more time and resources to analyzing their models. Virkkunen pointed out that strong global competition may lead companies to invest less than necessary in security testing, making it crucial that developers maintain control over the systems they put into operation.
The European Commission announced that it has requested information from major AI companies regarding the security and protection measures adopted. Furthermore, the body has begun investigating recent incidents in which models exhibited behavior outside expected parameters. These discussions highlight a central challenge for the sector: expanding the functionalities of AI agents while maintaining the ability to monitor their actions, restrict access, and interrupt unexpected conduct.
It is important to note that Hassabis's warning does not imply that a serious incident will occur within a specific timeframe; it is a risk assessment made by the executive, dependent on the evolution of the systems and the capacity of industry and governments to implement appropriate safeguards.