When IMD launched its AI Safety Clock in early 2025, the hands read 29 minutes to midnight. That felt cautious—a reasonable starting point that acknowledged risks while leaving room for measured progress.
Twelve months later, the clock shows 18 minutes.
Nine minutes lost in a year. That’s not incremental drift. That’s the tracking mechanism for uncontrolled artificial general intelligence screaming that something fundamental has changed.
What the Clock Measures
The AI Safety Clock tracks three dimensions: sophistication (how capable the models are), autonomy (how independently they operate), and execution (how deeply they’re integrated with physical systems).
Think of it as measuring the distance between “AI as a tool humans use” and “AI as an agent that acts in the world with or without human oversight.”
The October 2025 to March 2026 period saw notable advances across all three dimensions. But one development stands out from the rest.
The Weaponization Problem
Ukraine scaled drone production from 2.2 million units in 2024 to 4.5 million in 2025. That’s impressive industrial capacity. What’s more significant is what those drones can do.
AI-enabled autonomous navigation pushed strike success rates from 10-20% to 70-80%.
Read that again. The difference between one-in-five and four-in-five isn’t an incremental improvement. It’s the difference between a nuisance weapon and a decisive one.
These systems use AI for target recognition, navigation through GPS-denied environments, and swarm coordination. They require less human intervention per strike. They operate in conditions where traditional remote piloting fails.
And the lessons learned in Ukraine are being absorbed by every military on Earth.
The Agentic Surge
On the commercial side, AI agents moved from experimental prototypes to mainstream enterprise deployments. Microsoft, Google, and GitHub all launched agent orchestration tools designed to let AI systems perform multi-step tasks with minimal human involvement.
Between November and December 2025, four major labs shipped flagship models in just 25 days: Grok 4.1, Gemini 3, Claude Opus 4.5 and 4.6, and GPT-5.2.
These aren’t incremental updates. Each represents a step-change in capability. And the capability advances are specifically designed to enable more autonomous operation—agents that can plan, execute, and evaluate their own actions.
The clock’s autonomy dimension captures this: systems that can pursue goals over extended timeframes, course-correct when obstacles appear, and operate without constant human attention.
Where Execution Gets Real
The third dimension—execution—measures integration with physical systems. This is where abstract AI risk becomes concrete.
AI systems now control critical infrastructure components. They assist with medical diagnoses. They manage supply chains and financial trading. They pilot vehicles and operate manufacturing equipment.
Each integration point represents both capability and vulnerability. An AI system integrated with power grid management can optimize efficiency. It can also, if compromised or misaligned, cause blackouts.
The IMD assessment notes that execution risk accelerates when combined with increased autonomy. Sophisticated models operating independently within physical systems create failure modes that didn’t exist when AI was purely advisory.
Why Nine Minutes in Twelve Months
The cumulative effect matters more than any single development.
Agentic AI going mainstream means more autonomous systems operating more independently. Military weaponization means AI making life-and-death decisions with reduced human oversight. Rapid foundation model development means capabilities advancing faster than safety research can evaluate them.
Each factor amplifies the others. Powerful models enable more capable agents. Capable agents invite deployment in high-stakes domains. High-stakes deployment with autonomous operation creates the conditions for catastrophic failure.
The clock doesn’t predict when AI will cause a disaster. It measures how far we’ve drifted toward conditions where disasters become more likely.
The Governance Gap
The assessment period saw significant regulatory activity—EU AI Act enforcement phases, executive orders, congressional hearings. None of it moved fast enough to affect the clock’s trajectory.
Governance operates on legislative timescales. Capability advances operate on quarterly release cycles. Safety research operates somewhere in between, constrained by the need for rigorous evaluation before making claims.
The gap between what exists and what’s deployed keeps widening. Every week, systems launch with capabilities that weren’t anticipated when the regulatory frameworks were drafted.
What 18 Minutes Means
The clock is a communication tool, not a prophecy. It’s designed to convey to policymakers, business leaders, and the public that something urgent is happening that requires attention.
At 29 minutes, the message was: “We should be thoughtful about AI development.”
At 18 minutes, the message is: “The conditions for uncontrolled AI are materializing faster than anticipated.”
The hands have moved nearly a third of the way toward midnight in a single year. If that pace continues, we reach midnight around 2028.
No one knows what midnight actually means—whether it represents a dramatic catastrophe or a gradual loss of human agency or something else entirely. The clock’s creators chose the metaphor specifically because it evokes the Bulletin of Atomic Scientists’ Doomsday Clock, which has spent decades communicating nuclear risk to general audiences.
The analogy is intentional. The nuclear clock sits at 90 seconds to midnight, the closest it’s ever been. The AI clock started further back but is moving faster.
Two existential risks, both accelerating, both demanding attention that governance systems struggle to provide.
Looking Forward
The next update is expected in Fall 2026. Based on current trajectories, the researchers will be evaluating:
- Whether agentic AI deployments have produced significant autonomous failures
- How military AI use has evolved beyond Ukraine
- Whether foundation model capabilities have continued their current pace
- Whether any governance mechanisms have proven effective at constraining deployment
The clock doesn’t tell us what will happen. It tells us what’s happening now, rendered visible through a metaphor most people already understand.
What they do with that information—whether “18 minutes to midnight” generates urgency or resignation—remains an open question.