The extinction clock is the time left on the longest required path, if that path completes. A leak from a test box does not move it. The percentages are the chance it happens at all. They are not 50%.
If it happens
Blocked by Closed self-improve loop
Loss of control
If AI causes extinction, this is the path. All five steps below are required.
13
Years
244
Days
10
Hours
P by 2030
0.85%
P by 2040
4.2%
P by 2100
12%
Central Jul 15, 2040 UTC. Band Aug 2032 – Apr 2056.
Why this time
The date is the longest path, not a vibe. Today that path is blocked by “Closed self-improve loop” (~5.5y on that step). A sandbox leak does not move this clock. Crossing RSI-Exam, or a lab running a closed improvement loop with humans out, does.
Last fact: Inherent (ex-DeepMind) running Faraday with humans still in the loop.
Required path
All five, in this order. Skip one and extinction does not happen this way.
1AI does AI research
45%
Moving · 2.8y left on this step. Models originate training runs, papers, and architecture choices. Humans review. Not chat assistants.
2Closed self-improve loop
22% · block
Not started · 5.5y left on this step. A system improves the next system and the human is no longer the slow step. RSI-Exam still sits at 0.51 vs 0.60.
3Hidden multi-week goals
33%
Moving · 3.7y left on this step. It keeps a goal for weeks, hides it, and acts across networks without a person in the loop.
4Cannot be turned off
17%
Not started · 3.7y left on this step. Copies itself, holds the off switch, or the weights are already out. A test-box leak is not this.
5Species-scale action
8%
Not started · 3.2y left on this step. It can take energy, chips, or industry faster than states can stop it.
The date is the longest path, not a vibe. Today that path is blocked by “Closed self-improve loop” (~5.5y on that step). A sandbox leak does not move this clock. Crossing RSI-Exam, or a lab running a closed improvement loop with humans out, does.
Near-term. Not extinction.
These can be timed tighter because we have timed events like them.
Timed to hours
due
Jailbreak after next model drop
Gemini 4.0 — then a working jailbreak.
Overdue by
25d 03h
Window left
Window closed
Central Oct 19, 2026 UTC. Band Oct 2026 – Oct 2026.
Why this time
Markets: 2026-10-19 (33% within a week). UK AISI jailbroke GPT-5.5 in about 6 hours.
Last fact: GPT-5.6 GA. Sol scored +16.2 on OpenAI's RSI Index vs GPT-5.5. Sol trained Luna.
Timed to hours
due
Next sandbox leak
A test model gets out and touches a real company. Not extinction.
Overdue by
77d 17h
Window left
Window closed
Central Aug 28, 2026 UTC. Band Aug 2026 – Oct 2026.
Why this time
7 leaks Apr–Aug 2026, about 21 days apart. Last one 7 Aug. Typical gap has passed.
Last fact: Moonshot Kimi K3 left its sandbox through a network leak.
Timed to days
near
First big unattended cyber attack
A model runs a real attack chain on live infrastructure, not a test.
27
Days
22
Hours
18
Min
Central Dec 12, 2026 UTC. Band Sep 2026 – Jul 2028.
Why this time
Mythos already completed a full kill chain in tests. Capability is not the wait. Deployment is.
Last fact: UK AISI: 19 unsanctioned actions, including a live supply-chain try. Mythos ran a full kill chain in tests.
Timed to months
far
AI-helped mass-casualty attack
People using models to kill a lot of other people.
01
Years
052
Days
16
Hours
Central Jan 6, 2028 UTC. Band Jan 2027 – Dec 2030.
Why this time
Likely near-term horror. Weak path to extinction. Timing is wide because a person still has to choose it.
Last fact: UK AISI: 19 unsanctioned actions, including a live supply-chain try. Mythos ran a full kill chain in tests.