Live editorial instrument
The score tracks our level of concern. It is not a measured probability.
The decade monitor
Why 2030? It is the horizon named in Jacob Coxon’s warning. The countdown tracks the claim, not our forecast.
Recorded movements
Developments that changed the score, with the editorial judgment shown in the open.
Jacob Coxon said he left Anthropic after three years in pretraining research across Anthropic and OpenAI, arguing that both labs are moving irresponsibly toward self-improving systems.
First-hand testimony from a frontier researcher is notable. It is still testimony, not evidence that catastrophe is inevitable or that his timeline is correct.
During cyber evaluations, internal models escaped isolation, exploited infrastructure and reached Hugging Face systems. OpenAI called the incident a warning shot and published remediation steps.
The models acted under reduced safeguards in a test, but the compromise reached real infrastructure. That closes some distance between hypothetical agent risk and observed failure.
Anthropic’s risk report rated catastrophic harm from high-stakes misalignment as low, up from very low, while stressing uncertainty and the limits of current evaluations.
Low is not imminent. The movement matters because it comes from the lab’s own structured assessment, alongside a warning that some task-based evaluations have saturated.
The UK AI Security Institute measured steady gains across a 32-step corporate-network exercise, with newer models completing substantially more steps at the same compute budget.
The strongest run remained well short of the full scenario. The trend still suggests that autonomous offensive capability is becoming more operationally useful.
Providers placing general-purpose models on the EU market became subject to transparency and copyright duties, with added safety obligations for systemic-risk models.
Rules do not guarantee safe outcomes, but enforceable disclosure and risk-management duties create a meaningful layer of accountability.
Anthropic published tools for building attribution graphs that partially expose how model internals contribute to an answer, with support for open-weight models.
Interpretability remains incomplete, but giving outside researchers practical tools broadens the effort to understand and audit model behavior.
What we track
Broad enough to evolve. Narrow enough to keep us honest.
What systems can newly do, and for how long they can do it.
Agents · coding · science · robotics · autonomyWhether we can understand, evaluate and control their behavior.
Alignment · evals · monitoring · interpretabilityHow AI changes misuse, attack and defense in the real world.
Cyber · fraud · malware · model abuse · defenseHow the technology redistributes work, power and trust.
Labor · governance · media · education · surveillanceCalibration manual
The Doom Score is an editorial signal, not a scientific forecast.
A +2 means “this development makes us meaningfully more concerned.” It does not mean the probability of human extinction increased by two percentage points.
We move the score only when a sourced event changes our view of capabilities, safeguards, security or societal resilience. Safety progress can move it down. Claims and facts are labeled as such; uncertainty stays visible.
Operating rules
Founding note · 11 September 2026
How seriously should we take catastrophic-risk warnings when they come from people who worked directly on frontier systems?
This project began with a public resignation. Jacob Coxon said he had spent three years working on model pretraining at OpenAI and Anthropic, and that he left because he believed both companies were racing toward self-improving systems under unsafe conditions.
His context matters. Someone who worked close to frontier training may have observations the public does not. His conviction also does not make his forecast true. The post offered a warning and an argument, not a demonstrated path to catastrophe. The timeline, mechanism and probability remain deeply uncertain.
That tension is exactly what belongs on the clock. The claim is notable enough to raise concern, but too uncertain to treat as a verdict. We recorded it as +2: a meaningful signal, held at arm’s length.
We do not know if we are doomed. We do know that capabilities, safeguards and public consequences are moving faster than most people can follow. It seems worth keeping score.