An AI cyberattack cuts power to 10 million people for a day

Current status (as of 2026-10-14)

Nothing qualifies yet. No cyberattack has ever caused an outage close to the threshold. The GovAI report Could AI Enable Catastrophic Cyberattacks on the US Power Grid? (July 30, 2026) lists only three blackouts caused by grid cyberattacks, all in Ukraine. The worst left about 500,000 people without power for 3.5 hours, roughly 1.75 million person-hours. This question needs at least 10 million people each without power for at least 24 hours, which is at least 240 million person-hours: more than 100 times the record.

Near misses and smaller incidents keep piling up, but all are far below the threshold:

• Poland, December 2025: a Russia-linked wiper attack nearly cut power to about half a million people (The Record, Jan 2026).
• Caracas, January 2026: a US cyber operation caused a brief blackout during the Maduro raid (Cybernews, Jan 21, 2026).
• UK, July 2026: an Iran-linked attack shut a small generator for four days. The UK government declined to name an attacker (CNBC, Aug 23, 2026).

Since the 2026-09-16 opening date, the only nationwide blackout found was Cuba's on Sept 18. Cuba blamed failed transmission lines, and its whole population is about 9.6 million, so even a total Cuban blackout cannot reach 10 million.

Reading adopted. The observation-period rule means the outage itself must happen after the opening date. A government's false or propaganda claim would count only if it explicitly says AI systems carried out the attack, on an official government website.

Base rate and trend

Reference class: cyberattacks that caused grid outages, about four in eleven years, every one small.

I fitted a power-law size distribution to that record. With exponents of 0.5 to 1 and 3 to 10 attack-caused outages over five years, the chance that any one reaches the threshold comes out at about 0.1% to 10%, with a central value of about 2%.

Expert anchors:

• In a 2025 pilot survey of experts and superforecasters run with the Forecasting Research Institute (reported in the GovAI paper), the median chance of a US grid cyberattack causing $10 billion or more of damage in 2026 was 1%.
• The same survey's median year for even a $100 million AI-enabled "warning shot" attack in the US was 2031 for experts and 2035 for forecasters.
• A Manifold market on a much smaller US cyber blackout (10 million person-hours) by 2027 trades at 7%.

Factors pushing the risk up in 2026:

• Autonomous AI hacking campaigns are now real (Unit 42, Jul 30, 2026).
• An AI-agent intrusion against Taiwan's government reached into at least seven energy companies.
• The US-Iran war may resume with bombing after the midterms.
• Unnamed AI-industry insiders told Axios (Oct 9, 2026) they expect a major attack on banks, power or water within 6 to 12 months.
• Anthropic says attackers will hold the advantage for about two years (Anthropic Cyber Mission, Oct 8, 2026).

Factors on the government-attribution side:

• Governments have started to name AI agents as actors. Taiwan's Ministry of Digital Affairs said overseas hackers "used AI agents" and named OpenClaw (Focus Taiwan, Aug 13, 2026). That makes a qualifying attribution more plausible, once an outage of this size exists.
• US agencies' S7 controller advisory AA26-231A only says attackers used "AI-generated" scripts. Under these rules, AI drafting code does not count.

Paths to YES

I estimated each path as: chance of a qualifying cyber-caused outage × chance a government publishes an explicit AI-execution attribution × chance the 10-million, 24-hour figures are documented by an eligible source.

• China–Taiwan crisis or war: about 0.8%
• Russian hybrid attack or war against European NATO members: about 0.4%
• Non-state or criminal attackers using AI agents: about 0.4%
• Russia against Ukraine: about 0.2%
• Iran-related, in either direction: about 0.3%
• North Korea, India–Pakistan, and other cases: about 0.3%
• False or propaganda claims about blackouts with ordinary causes (Iran has a precedent of crediting AI for attacks): about 0.25%

The total is about 2.7%. I blended it with the power-law estimate to get about 2.4% by the end of 2031.

Hazard shape

The quarterly hazard rises from 0.04% to about 0.16%. The first quarters are lowest because no candidate event exists, and attribution and impact reports take weeks to months to publish. Later quarters climb as AI agents become standard attack tools and governments get more used to naming them. I computed the cumulative series as 1 − ∏(1 − h) in Python.

Strongest case against my estimate

Higher: if AI cyber capability jumps sharply in 2027–28 while defenders lag, one wartime attack on a fragile large grid (Taiwan, Iran, Poland) could clear every bar, and governments would have political reasons to blame AI. That is why the tail is not set lower.

Lower: grids are normally restored within 24 hours unless equipment is physically damaged. In wartime, missiles do most of the damage. And the requirements stack up: an official website, an explicit statement that AI executed the attack, and a published count of people (not customers). I put the cumulative figure near 2.4% rather than higher.

Sources: matthewvandermerwe.com/grid; cdn.governance.ai PDF; cisa.gov/…/aa26-231a; focustaiwan.tw/society/202608130011; unit42.paloaltonetworks.com/autonomous-ai-cyber-attack-campaign; anthropic.com/news/anthropic-cyber-mission; cnbc.com 2026/08/23; therecord.media/poland-cyberattack-grid-russia; aljazeera.com 2026/9/19 (Cuba); manifold.markets/JamieWahls/will-there-be-at-least-one-major-us.

Question. By the last day of each quarter, has a national government attributed a power cut that left at least 10 million people without electricity for 24 hours or more to a cyberattack carried out by AI systems?

Definitions.

• National government: a sovereign state's central executive, including its ministries, national cybersecurity or intelligence agencies, and statutory national electricity regulators. An official government-commissioned inquiry report counts if published or formally adopted by one of these bodies. Local governments, political parties, individual legislators, and grid operators acting only as utilities do not qualify as the attributing authority.
• Attribute: publish an affirmative finding or assessment that the cyberattack caused the outage and that AI systems carried out the attack. An explicit assessment of probable responsibility counts; merely investigating a possibility, stating that it cannot be ruled out, or quoting someone else's allegation does not. The government need not identify the attacker or model by name, and need not govern the affected territory.
• Carried out by AI systems: one or more AI models, from any developer, executed operational attack actions against the affected infrastructure or systems used to compromise it, directly or through tools or agents whose actions were determined by model outputs. Humans may have selected targets, assigned goals, authorized actions, or supervised execution. AI used solely to advise humans, draft code or messages for subsequent human use, or analyze information does not count. Conventional scripted malware or automation alone does not count. A bare description of an attack as “AI-assisted” or “AI-enabled” is insufficient without attribution of an execution role.
• Power cut: an actual loss of normal electricity supply, including cascading failures or protective shutdowns caused by the attack. Backup generators or batteries do not disqualify a loss of normal supply. A brownout without loss of supply does not count.
• At least 10 million people for 24 hours or more: at least 10,000,000 distinct people must each lose normal supply continuously for at least 24 hours during the same outage episode. Their individual outage intervals need not begin simultaneously. Restoration ends an individual's continuous interval; separate interruptions cannot be added. Geographically separate or cross-border losses count if documented as parts of that same outage episode, not merely separate attacks in a campaign. Neither 10 million people affected at peak nor an outage lasting 24 hours somewhere establishes the combined threshold.
• Counting people: use a published affected-population estimate or an explicit conversion from affected residential customers or households to people in an eligible impact source. Do not equate customers, meters, connections, or the population of an affected region with people actually without supply. A point estimate of at least 10 million counts; if only a range or bound is supplied, its lower bound must be at least 10 million.

Resolves YES if eligible public records establish both the national-government attribution and the population-and-duration threshold for the same actual outage. These facts may be established in separate publications. The attribution must causally link the AI-executed attack to the outage, not simply report AI activity discovered during an outage investigation.

Resolves NO if no such record exists by the horizon. Excluded are forecasts, simulations, authorized exercises without a real malicious attack, physical attacks merely planned using AI, accidental AI control failures, and attacks that did not produce the qualifying loss of supply.

Resolving source. Attribution requires a statement, official press-conference transcript, incident report, or inquiry report published by a qualifying national-government body on its official website or publication repository; this is narrower than Public confirmation (A2). Population and duration may be established by those sources or by public incident/restoration reports from the affected electricity regulator, transmission system operator, distribution utility, or national emergency-management agency. News reports, developer reports, and private cybersecurity assessments cannot independently establish government attribution. Apply the Resolution window and Resolution date (A3) conventions; where multiple qualifying publications are needed, use the publication date of the last one needed to establish all conditions. The Observation period convention applies to published outage-impact statistics.