Anthropic and OpenAI Are Rehearsing Their Response to the First Major AI Catastrophe

Leaders at Anthropic, OpenAI and several other AI companies are privately running through scenarios for the first large-scale AI-related incident. They focus above all on the outside reaction: how the public and politicians would respond once a failure affects millions of people. One scenario is a cyberattack that leaves people without access to financial services and the internet and, in the worst case, without electricity and water.

Few details are public. The source gives no figures, probabilities or timelines, and does not say which measures the companies are preparing. But the very fact of such drills says a lot. For years the industry discussed risks mostly in the language of technical reports. Now it is rehearsing a scenario in which the explanations will go to lawmakers and voters, not researchers.

What is known so far

• Participants: leaders at Anthropic, OpenAI and other AI companies.
• Format: closed-door discussions, scenario "play-throughs."
• Subject: public anger and the political reaction to a major incident.
• Key scenario: a cyberattack on financial services, the internet, the power grid and water supply.
• Unknown: probability estimates, possible protective measures, the government's role in these plans.

Why a cyberattack is the focus

The choice of scenario is no accident. Of all the risks discussed around advanced models, cybersecurity is the most down-to-earth. It requires no hypothetical superhuman abilities: a model that writes code, finds vulnerabilities and acts as an autonomous agent is enough. Both labs have already built this risk into their own frameworks. At OpenAI, cybersecurity is one of the categories in the Preparedness Framework; Anthropic has a Responsible Scaling Policy with safety levels. In November 2025, Anthropic itself reported a campaign in which attackers used Claude Code to automate a significant part of an espionage operation. It was not a catastrophic incident, but it showed where practice is heading.

Critical infrastructure adds a social problem to the technical one. Bank payments, communications, power grids and water utilities are interconnected. A failure in one link quickly turns into a cascade, and people left without money, connectivity or water will not care who is to blame: the attacker, the model developer or the network operator.

The political blow matters more than the technical one

The history of technology crises shows that regulation after a high-profile disaster arrives fast and is rarely careful. If an incident is linked to AI, the whole industry will come under pressure, including companies that had nothing to do with it. Possible consequences include strict requirements on access to models, mandatory reporting and liability for developers. These are logical assumptions, not announced plans of the companies.

That is why the labs are preparing not only technical "Plan Bs" but also communications and relationships with regulators. Whoever explains what happened first, and shows what they did in advance, will earn more trust. Whoever stays silent risks getting a law written in a hurry.

What teams and businesses should do

The labs' drills do not replace your own preparation. If your product depends on an AI provider, check a few things:

• Map your dependencies: which processes would stop if the API were unavailable for several hours or days.
• Prepare a backup provider or a local model for critical functions.
• Limit the permissions of AI agents: minimal access to payment, production and infrastructure systems, and mandatory human confirmation of dangerous actions.
• Add an "AI component compromised or unavailable" scenario to your incident response plan and practice it in drills.
• Keep manual procedures for payments and communications wherever a stoppage hurts people.

The main takeaway is simple: the industry itself treats a major incident as a question of "when," not "if." It is wiser to prepare in advance, while the scenario is still a drill.