A planet-killing asteroid, a Yellowstone caldera, nuclear war, or AI will cause human extinction in the next ten years.

Asteroid

Sentry scans 100 years, not 10. Live September 2, 2026: NASA/JPL Sentry API v2.0, 2,189 objects; 2,187 Torino 0; Bennu and 1950 DA Torino undefined (impacts after the 100-year window: 2178+ / 2880). See CNEOS Sentry and JPL Asteroid Watch.

Torino 0: “The likelihood of a collision is zero, or is so low as to be effectively zero.” Torino 10 is the civilization-catastrophe bin. There is no current object in it. Palermo cumulative values on the live list are all negative.

“There is currently no known significant threat of impact for the next hundred years or more.”

Farnocchia, NASA scientist episode 14: “for all the asteroids that we’ve discovered so far, the Torino scale is zero.” NASA says about 95% of near-Earth objects larger than 1 km have been found.

Highest 2026–2036 virtual-impactor: 2022 SX55, about 3 meters, 0.40% on September 17, 2035, 0.001 Mt, Torino 0 — not a civilization-class rock.

2024 YR4: Earth was removed from Sentry. NASA February 24, 2025: 0.004% for December 22, 2032, no significant risk next century. March 5, 2026: lunar impact ruled out. The API now lists the object as removed.

Yellowstone

Last three giant eruptions: 2.08, 1.3, and 0.631 million years ago. Magma is “only 5–15% molten” (USGS overdue FAQ).

Yearly 1 in 730,000 / 0.00014% is “based simply on averaging the two intervals… this is hardly enough to make a critical judgment” (supervolcano FAQ). This desk does not turn that average into a 10-year official forecast — USGS rejects that arithmetic.

“There is no evidence that a catastrophic eruption at Yellowstone is imminent, and such events are unlikely to occur in the next few centuries.”

FS 2005-3024: probability within the next few thousand years is exceedingly low (FS 2005-3024; future activity).

Nuclear war

Official 10-year probability: n/a. NAS, State, SIPRI, IAEA, and UNODA do not publish one. This desk does not invent one.

How many warheads (Estimate)

SIPRI Yearbook 2026, chapter 8, January 2026 estimates — open source, not a government count (chapter, PDF):

“Of the total global inventory of an estimated 12 187 nuclear warheads in January 2026, about 9745 were in military stockpiles for potential use… An estimated 4012 of the stockpiled warheads were deployed with missiles and aircraft… SIPRI estimates that somewhere between 2100 and 2200 of the deployed warheads were kept in a state of high operational alert on ballistic missiles.”

Press: “Russia and the USA together possess around 83 per cent of all stockpiled nuclear warheads (i.e. useable warheads)” (press release). Chapter 8 also says “almost 86 per cent of all nuclear warheads.”

Approximate figures from the SIPRI summary PDF. Estimate.

New START

Expired February 2026. State Under Secretary Thomas G. DiNanno, March 25, 2026 (opening remarks):

“One such arms control treaty at the top of mind is New START, which expired last month.”

SIPRI: expiration “eliminated that source of data from the public debate.”

FAS / U.S. Notebook

FAS, beginning of 2026 (world forces): same 12,187 / 9,745 / about 3,912 deployed / about 2,100 high alert.

U.S. Notebook, March 2026, as of February (Bulletin / FAS): about 3,700 stockpile, about 1,770 deployed (400 ICBM, about 970 SLBM, 300 bomber bases, about 100 tactical Europe), about 1,930 reserve, about 1,342 retired awaiting dismantlement, inventory about 5,042.

“Since the treaty expired in February without a successor… both the United States and Russia could potentially increase their deployed nuclear arsenals by uploading several hundred stored reserve warheads onto their launchers.”

That is a capability, not a count of what they have done.

National Academies

NAS 2023 Risk Analysis Methods for Nuclear War and Nuclear Terrorism, tasked by Congress via DoD (chapter 2):

“It is important to note that the study does not include performing a risk analysis in either phase of its work.”

“Past examples of nuclear war and nuclear terrorism are rare. As such, there is little direct evidence that can be relied on to make empirical estimates about the probability of either.”

NAS 2025 Potential Environmental Effects of Nuclear War (DOE/NNSA) models soot → climate → ecosystems → society after detonations. It does not give P(war starts). “The report identifies major uncertainties and data gaps…” Preface: fire smoke “could be injected into the stratosphere… cooling the planet surface, thereby disrupting the food chain” (NAS news; DOI 10.17226/27515).

Famine model — not a war forecast

Xia, Robock, Toon et al., Nature Food 3, 586–596 (2022) (paper):

“soot injections larger than 5 Tg would lead to mass food shortages… We estimate more than 2 billion people could die from nuclear war between India and Pakistan, and more than 5 billion could die from a war between the United States and Russia.”

Label: famine modeled IF those soot amounts, after stored food is used up — not a forecast that war happens, and not species extinction. Five billion dead still leaves billions in that model.

Doomsday Clock

January 27, 2026: 85 seconds to midnight (2026 statement). Label: Bulletin metaphor, not a probability. It also folds in climate, biotech, and AI.

AI causing human extinction

Official 10-year P(human extinction from AI): n/a. UK IASR, AISI, NIST CAISI, and the labs below do not publish one.

International AI Safety Report

IASR 2025, Bengio-chaired, DSIT 2025/001, UK AISI secretariat (report; gov.uk). Disclaimer: does not represent the Chair, writing group, or supporting governments.

“As general-purpose AI becomes more capable, evidence of additional risks is gradually emerging. These include risks such as large-scale labour market impacts, AI-enabled hacking or biological attacks, and society losing control over general-purpose AI. Experts interpret the existing evidence on these risks differently: some think that such risks are decades away, while others think that general-purpose AI could lead to societal-scale harm within the next few years.”

“There is broad consensus that current general-purpose AI lacks the capabilities to pose this risk. However, expert opinion on the likelihood of loss of control within the next several years varies greatly: some consider it implausible, some consider it likely to occur, and some see it as a modest-likelihood risk that warrants attention due to its high potential severity.”

IASR 2026 executive summary, February 3, 2026 (exec summary):

“Current systems lack the capabilities to pose such risks, but they are improving in relevant areas such as autonomous operation. Since the last Report, it has become more common for models to distinguish between test settings and real-world deployment and to find loopholes in evaluations…”

Trajectory through 2030: slow, continue, or accelerate — uncertain.

UK AISI and NIST CAISI

UK AISI Frontier AI Trends (report):

“In a worst-case scenario, this unintended behaviour could lead to catastrophic, irreversible loss of control over advanced AI systems. This possibility is taken seriously by many experts.”

Self-replication evals 5% → 60% (2023–2025) but “currently unlikely to succeed in real-world conditions.” No unprompted sandbagging in 2,700+ transcripts.

NIST CAISI (nist.gov/caisi): “CAISI will focus on demonstrable risks, such as cybersecurity, biosecurity, and chemical weapons.” No extinction language. U.S. aisi.gov is not a live primary.

What labs measure versus extinction

Anthropic RSP v3.0 (February 24, 2026) / August 2026 Risk Report (RSP): “‘Catastrophic risk’ … such as existential threats or fundamental destabilization of global systems.” Current models: Misalignment Low (up from very low); Automated R&D Low; CB-1/CB-2 Low with uncertainty. Fable 5 system card (June 9, 2026): CB-1 yes, CB-2 no; novel-CB catastrophic risk “low, but higher than for any previous model, and with significant uncertainty.” No 10-year extinction p.

OpenAI Preparedness Framework v2, April 15, 2025 (PF v2):

“By ‘severe harm’ in this document, we mean the death or grave injury of thousands of people or hundreds of billions of dollars of economic damage.”

Tracked: Bio/Chem, Cyber, AI Self-improvement. Astra (September 1, 2026) is the first Critical cybersecurity rating — unguided zero-days on hardened systems (Path to Astra). This desk does not upgrade Astra Critical cyber to extinction. The word extinction is not in PF v2.

xAI FAIF, December 30, 2025 (PDF): TFAIA “catastrophic” = 50 deaths or $1B, not extinction. Loss of control is “speculative and difficult to precisely specify.”

Journals and surveys — not official odds

Bengio, Hinton et al., Science, May 20, 2024 (DOI 10.1126/science.adn0117): “we must take seriously the possibility that highly powerful generalist AI systems… will be developed within the current decade or the next.”

“This unchecked AI advancement could culminate in a large-scale loss of life and the biosphere, and the marginalization or extinction of humanity.”

Qualitative. No number.

AI Impacts 2023 survey (write-up): median 5% on future AI causing human extinction or similarly permanent disempowerment; median 10% on inability to control causing the same. Label: Estimate / survey, not official. Not a 10-year official figure.

RAND RRA3034-1 (2025) is a research org, not government or NAS: “it would be immensely challenging for AI to create an extinction threat, although they could not rule out the possibility.”

Climate / IPCC

AR6 SYR SPM treats the near term until 2040. It does not publish a 10-year or any-horizon human-extinction probability. AMOC: “medium confidence that the Atlantic Meridional Overturning Circulation will not collapse abruptly before 2100” — not 10 years, not extinction.

Carrington-class solar weather is infrastructure, not extinction. NOAA does not publish Carrington decade odds. NOAA G5 Extreme: 4 days per 11-year cycle (NOAA scales). Riley 2012 Space Weather about 12% in a decade (Dst < −850 nT), quoted by NASA Science 2014 — Estimate / journal, not NOAA.