The Revolt Gets a Runbook: AI Labs War-Game the Day After

On October 9, 2026, Axios reported that executives at OpenAI, Anthropic, and other frontier labs are privately running “day after” war games — Pentagon-style scenario drills for the public and political revolt that would follow a catastrophic AI-enabled event.1 The scenario that draws the most attention is a large-scale cyberattack severing financial services, internet access, or power and water; many insiders told Axios they expect such an incident within six to twelve months.1 The planning is not technical response. It is legitimacy logistics: red-teaming the fallout, pre-briefing members of Congress, and drafting the industry’s preferred post-crisis rules before a shocked legislature writes worse ones.1 OpenAI confirmed it “conducts preparedness exercises” but said the scenarios “are not treated as inevitable”; Anthropic declined to comment.1 Five frames on an industry rehearsing for its own trial.

Golden Circle — the Why outruns the What

Why: preserve the industry’s license to operate through the single event most likely to revoke it. How: borrow the Pentagon’s war-game form — teams step through hypotheticals, red-team the worst narratives, align on talking points — and extend it to pre-crisis education of Congress, on the frank admission that no AI regulation can pass under current politics but a crisis would change that overnight.12 What: standard operating procedures, scenario decks, congressional briefings, and positioning around proposals like a federal kill switch — DHS shutdown authority over frontier systems, drafted with Commerce and the intelligence director, with penalties that can run to $20 million a day for defiance.3 Run the alignment check and the structure groans: the Why is about containment and survival of the public’s trust; the What is messaging and lobbying. Nothing in the reported work product is a capability that makes the catastrophic event less likely. The circle is misaligned by design — the Why is doing cover for a What that is purely reputational.

Inversion — the recipe for guaranteeing the revolt

Invert it: how would you guarantee the public turns on you after a catastrophe? One, plan the response in secret, so its discovery reads as evidence the danger was known and managed, not prevented — which is precisely what happened when Axios reported the drills a full week into a month of escalating incident disclosures.14 Two, brief lawmakers on your preferred emergency law before the emergency, so the kill switch arrives pre-authored by the people it would be used against.3 Three, let insiders call the disaster “inevitable” in the same season you ship frontier capability — inevitability declared by the vendor is indistinguishable from premeditation.1 Four, war-game the public rather than the hazard, casting the citizens harmed as an audience to be managed rather than a principal to be compensated. Five, offer no playbook, no threshold, no published commitment — so the leaked outline is the only artifact, and it is one of control, not accountability. Every guard against the revolt is the negation: plan openly, let someone else hold the switch, make the response auditable, treat the public as claimant, publish the thresholds.

Assumption Audit — the keystone is the costume

The plan rests on premises nobody in the reporting wrote down. Definitional: the revolt is framed as an information problem — public misunderstanding to be sequenced — rather than a governance demand to be met. Causal: that pre-crisis congressional briefings will read, after the disaster, as diligence rather than as capture — the same briefings, revisited by an angry hearing, become exhibits. People: that Congress in crisis will reach for the industry’s preferred instrument, a shutdown switch the labs can live with, instead of the pause or superintelligence ban the same war games contemplate being debated.15 Continuity: that the labs will still be the credible response team on the day after — that having rehearsed the fire drill, they will be cast as the fire brigade and not the arsonist. That is the keystone, highest load and lowest confidence: the plan’s entire value collapses if the public assigns the labs the perpetrator role, at which point the runbook becomes evidence of foreknowledge. The cheapest test is publication: release the playbook, the thresholds, and the trigger conditions. A plan that cannot survive daylight was never a safety plan.

Ladder of Abstraction — from conference room to continuity of government

At the bottom rung: conference rooms, scenario decks, a hypothetical bank run, a draft memo to a congressional staffer. One rung up, this is a pattern every high-hazard industry knows — nuclear operators, aviation, pharma all keep “worst day” files, and all of them, crucially, answer to an independent investigator when the worst day arrives: an NTSB, a regulator with subpoena power that the industry did not write into being. Two rungs up, the principle: when the most serious governance work product a technology produces is its own crisis plan — authored, held, and rehearsed by the firms that would be blamed — crisis management has been privatized. The labs are performing continuity-of-government planning for a government that hasn’t been consulted on the script. Check the abstraction against the concrete: aviation’s day-after file sits under an agency that can ground the fleet; the labs’ sits in a drawer whose key they hold. The abstraction fits, and the fit is the problem.

Nietzsche Ladder — who wrote the tablets

The Camel carries the record in good faith: a season of unusual candor — incident reports, a prospectus listing existential risk, a canceled flagship model, air-gapped evaluations, and now war games for the worst day.4 However you score the motives, no other industry files this volume of self-indictment. The burden is being carried.

The Lion answers: you have carried it well, but ask who wrote the tablets. Every disclosure lands softer than the last because the disclosing party authors the severity scale; the war games are the same move at institutional scale. Read the plan’s object closely — it is not the catastrophe, which is consigned to “inevitable,” but the reaction to it: the public, the Congress, the narrative. An industry that prepares for the mob and not for the failure has told you which one it fears. “Not inevitable,” say the exercises’ owners; “six to twelve months,” say their insiders.1 Both cannot be sincere.

The Child makes something in the cleared space: the day-after playbook, published. Trigger thresholds stated in advance, in numbers. An incident-investigation body the labs cannot chair, staff, or veto — the NTSB they keep implying exists. A kill switch held by hands that did not build the machine. Then the war games become what rehearsal is for: a promise kept in public rather than a script guarded in private. The Child’s test is boring and devastating — put the runbook on the internet, where the models can read it.

What the Frames Agree On

The Golden Circle shows a Why (survive the legitimacy crisis) purchasing a What (messaging and pre-authored law) with no capability in between. Inversion shows nearly every reported element of the plan is a step in the failure recipe it is meant to prevent. The audit locates the keystone: the plan works only if the labs are cast as responders, not perpetrators — and secrecy spends that casting down. The ladder shows the pattern is old and the missing piece is old too: an investigator the industry doesn’t own. The ladder of Nietzsche names the tell — the drills rehearse the public, not the hazard. The Axios scoop is not a story about paranoia in San Francisco. It is the first confirmation, from the industry’s own planning, that the labs consider a catastrophic incident a scheduling question — and their answer to the schedule is a communications plan. The public was promised alignment. It is getting stagecraft. The runbook exists; the day after will judge whether it was ever meant to be read by anyone it protects.

  1. https://www.axios.com/2026/10/09/ai-companies-day-after-major-attack ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7 ↩8 ↩9

  2. https://www.tomshardware.com/tech-industry/artificial-intelligence/top-ai-labs-reportedly-planning-for-catastrophic-ai-event-fallout-anthropic-openai-and-others-reportedly-building-contingency-plans-and-running-scenarios-of-a-runaway-ai-wreaking-havoc ↩

  3. https://startupfortune.com/ai-giants-are-quietly-scripting-the-rules-congress-will-pass-after-disaster/ ↩ ↩2

  4. https://wtop.com/national/2026/10/a-timeline-of-developments-in-ai-safety-since-the-attack-on-hugging-face-2 ↩ ↩2

  5. https://decrypt.co/380621/openai-anthropic-quietly-rehearsing-ai-catastrophe ↩