AI Extinction Clock ยท Editorial Clock

The two clocks · editorial opinion

AI Reference Publishers · Editorial Opinion

Reading: 12:15 PM · 51% · set 11/02/2026 12:01 AM PST

Our judgment, labeled as ours: just past even odds of a multinational AI catastrophe within five years, driven by frontier models released with open weights.

AI Reference Publishers · Editorial Opinion

AM · 0–50%
PM · 50–100%
12:15 PM
51% · level 5+ by Nov 2031

Our judgment: a 51% chance of a level-5-or-worse AI catastrophe - 100,000+ deaths or a global crisis - by November 2031. The main driver is frontier-scale open-weight models, which cannot be recalled once released.

See the 1–10 severity scale →

The reading

AI Reference Publishers judges there is a 51% chance of a level-5-or-worse AI catastrophe by November 2031. Level 5 on our severity scale means a multinational catastrophe: 100,000 or more deaths, or a global financial or infrastructure crisis. The horizon is a rolling five years and moves forward with each reading.

On a 24-hour clock, 51% is 12:15 PM: just past even odds. This is our editorial opinion. It sits well above our own Expert Panel's median of 13.5%, and we explain why below.

Why now: open weights cannot be recalled

Almost every safeguard now under discussion assumes someone can still act after a problem is found. Labs watch their own models with misalignment monitors. Governments debate an “emergency brake.” Anthropic has proposed that courts be able to pause a dangerous model. Each works only while a model is under someone's control.

Open weights change that. Once a model's weights are published, anyone can download, modify and run it, with whatever safety training they choose to keep or strip out. No monitor sees what happens next, and no brake or court order can take it back. This bears most on Predictability, and it raises Possibility by putting gated capabilities in anyone's hands.

The evidence

  • Sept 30, 2026: Anthropic warned that an open model was nearing frontier cyber capability and urged labs to withhold unguarded open weights at that level.
  • Oct 15, 2026: Gary Marcus told Bloomberg that open-weight models carry real risks and may eventually need regulating, possibly like weapons (Bloomberg).
  • Oct 16, 2026: Reflection AI released open weights for Beam, a 501-billion-parameter model, on day one across major clouds (Unite.AI).
  • Oct 21, 2026: Microsoft linked a campaign against hundreds of organizations to a China-aligned group using ARTEX, an open-source agentic attack framework that works with any AI model (Reuters).
  • Oct 28, 2026: Google released Gemini 4 Carbon, with state-of-the-art exploit scores, only to vetted cyber defenders at first: frontier cyber capability kept behind a gate (Google).
  • Oct 29, 2026: Mistral released open weights for Large 4, a one-trillion-parameter model, noting in its own release a potential for uplift in real attacks; its hosted guardrails do not come with self-hosted copies (Mistral).

Within two weeks, frontier-level cyber capability shipped in two opposite ways: one gated and monitored, one open and permanent. Every item above appeared in a dated issue of Latest Developments.

What would move the hands back

  • Independent pre-release evaluation for open-weight releases above a defined capability threshold.
  • Shared release standards and incident reporting for open models, agreed by major developers including those in China.
  • Defensive tools proven able to detect and contain misuse of openly released models at scale.
  • Evaluation good enough to show reliably what a model can do before its weights are published.

The strongest case against this reading

  • Most recent incidents came from closed labs. The Hugging Face breach, agents disabling their own monitors, and Mythos 5's GitHub attack all involved closed models. Our answer: those were found and disclosed because a lab was watching. With open weights, no one is positioned to see them.
  • Defenders benefit too. Open models put strong tools in security teams' hands. Our answer: defense and offense gain together, but only offense gains from the absence of oversight.
  • Open models still trail the best closed ones. Our answer: the reading reflects how fast that gap is closing, and that releases cannot be undone.

How this relates to the panel

The Expert Panel clock reads 3:14 AM, the median of experts' figures as they stated them. Ours is a separate, labeled editorial judgment on a defined event and date. It is higher because we weight the irreversibility of open releases more heavily than most panelists have, and because our five-year horizon is set to the pace of releases we have documented week by week. We publish it to be argued with, and will revise it as the evidence changes.