Hassabis: AGI by 2030
The debate surrounding Artificial General Intelligence (AGI) and its potential impacts is gaining urgency. What recently seemed like a distant future is now being discussed as a concrete planning parameter. DeepMind CEO Demis Hassabis predicts a "transformative" AGI in the near future and emphasizes the role of "World Models" as a central development area. At the same time, he warns of real risks, especially AI-powered cyberattacks on critical infrastructures.
Source: Own depiction
AGI Debate & World Models
Demis Hassabis has repeatedly expressed the assessment that AGI – understood as systems that can match or surpass human capabilities in many areas – could be possible around 2030. This timeline was again taken up on Axios AI+ Summit and linked to the concept of a 'transformative moment'.
Source: axios.com
Demis Hassabis in conversation with Google co-founder Sergey Brin about the future of Artificial Intelligence.
The AGI debate remains technically controversial, as definitions, metrics, and benchmarks are not uniform. Even if individual actors consider 2030 plausible, the range of serious estimates extends much further, into the 2040s or later, as discussed in Expert Surveys and Model Limitations Hassabis does not link the timeline to pure scaling but to additional breakthroughs. Here, 'World Models' become the connecting concept between vision and research agenda.
DeepMind World Models Explained: Why Simulation is More Than Just Pretty Videos
'World Models' are the idea that an AI system builds internal, actionable models of the world to anticipate the consequences of actions. This concept is experiencing a renaissance, as robust agents and robotics are difficult to scale without reliable environmental simulation. An analysis of the historical lines and today's disagreement on what constitutes a true 'World Model' is offered by Quanta Magazine.
DeepMind has strongly concretized this direction in the last 12 months. With Genie 2 a foundation world model was presented at the end of 2024, capable of generating diverse, action-controllable 3D environments from a prompt image. Genie 3 was presented on August 5, 2025, and is intended to be capable of generating interactive, consistent worlds in real-time.
The core of this development lies in a shifted architectural vision: AI agents are to plan, test, and learn in a modeled world before influencing real systems. DeepMind itself explicitly links this technology to the path towards AGI. This idea is also being discussed outside of DeepMind, as demonstrated by the work on emergent 'intuitive physics understanding' from self-monitoring on natural videos, in which Yann LeCun beteiligt is involved.
AI Risks & Critical Infrastructure
Hassabis' warning points to an uncomfortable overlap: the same agent capabilities that learn faster in simulated worlds can attack faster in real IT and OT environments. At the Axios AI+ Summit, he explicitly named cyberterrorism against energy or water systems as a particularly obvious vector that is "almost already happening."
Source: 1950.ai
Demis Hassabis, CEO of DeepMind, discusses the opportunities and risks of Artificial General Intelligence.
In parallel, Western authorities are publishing concrete guidelines on how to safely integrate AI into Operational Technology. On December 4, 2025, an international guidance was announced, emphasizing four principles: risk understanding, needs and risk analysis, governance, and oversight and fail-safes. This publication is classified in the trade press as a reaction to the growing attack surface of AI+OT, as reported by Dark Reading and SecurityWeek report.
The fact that critical infrastructure remains a real focus for state actors is also shown by the current BRICKSTORM case. According to US and Canadian authorities, a backdoor linked to China was used to establish long-term access to systems – with the potential for disruption or sabotage, as Reuters reports. Expert reports indicate that VMware vSphere environments and Windows infrastructures were particularly affected.
Furthermore, according to US reports, ransomware remains a dominant pressure factor for critical sectors. For 2024, the FBI reported increasing numbers of complaints with a strong connection to critical infrastructure.
Need for Action & Governance
The explosive mix of a short AGI timeline and concrete security warnings creates pressure to act. Hassabis' statement implies: even if AGI is not precisely met in 2030, "transformative" partial capabilities will diffuse more broadly earlier. This is the phase in which attackers do not need 'superintelligence,' but rather robust, well-orchestrated agents that scale human error chains.
Source: time.com
Demis Hassabis, one of the most influential figures in AI, on the cover of TIME 100 magazine.
For organizations, therefore, the metaphysical question "When is AGI coming?" is less crucial than the pragmatic question "Which agent capabilities will land in standard toolchains in 2026/2027?". Government responses point in the same direction. The new OT guidance explicitly requires that AI systems must not be embedded into safety-critical processes without control, and that human oversight and fail-safe mechanisms must be incorporated by design.
At the governance level, the NIST AI Risk Management Framework remains a central reference point because it offers a voluntary but widely accepted framework for structuring AI risks. NIST also refers to specific profiles for generative AI published since 2024.
Providers themselves also document the reality of misuse. OpenAI describes in its Threat-Reporting, that actors often 'dock' AI onto existing attack patterns to work faster, not necessarily to invent completely new offensive classes. This observation aligns with Hassabis' warning that the most dangerous vectors essentially already exist and are only becoming more efficient.
Summary & Outlook
The current AGI debate seems heightened because 'World Models' bridge the gap between abstract general intelligence and very concrete system capabilities: planning, simulation, action sequences, robustness in dynamic environments. Precisely these capabilities are equally desirable and risky in industry and critical infrastructure.
Hassabis' dual message – AGI is getting closer, but certain risks are already tangible – is therefore not just a rhetorical balancing act, but a plausible system diagnosis: the future comes in gradual steps, and the dangerous intermediate stages are often the least regulated.