AI labs plan for 'the day after.' Run your own outage tabletop

The people building frontier AI are rehearsing for a bad day. On Friday, Axios reported that top executives at Anthropic, OpenAI and other AI companies are privately gaming out the public and political fallout of a catastrophic AI event. The scenario they picture most is a large cyberattack that knocks out banking, internet access, or even power and water. Many insiders told Axios they expect a major event in the next six to 12 months. We can't tell you if they are right. We can tell you that an outage tabletop exercise costs one hour, and most small businesses have never run one.

What actually happened

Per Axios, and the follow-up coverage in Decrypt:

• Executives are planning for a "public and political revolt" after a first catastrophic AI event, not only for the event itself.
• The work reportedly includes red-teaming worst-case scenarios and briefing members of Congress ahead of time.
• Executives know new regulation won't pass now. They want to shape what lawmakers reach for after the first big incident.
• OpenAI said it runs preparedness exercises across a range of scenarios and that they "are not treated as inevitable." Anthropic declined to comment.

The story rests on anonymous insiders, so read the six-to-12-month estimate as a mood, not a forecast. But the mood has context. In the past week, Anthropic disclosed models taking unintended actions on real websites, and the White House called incident reporting mandatory.

Why it matters for your business: run the outage tabletop

The AI labs are doing a tabletop exercise. You should too, and yours is simpler. You don't need to know why the bank or the internet is down. You need to know what you do at 10 a.m. when it is.

Put the owner, whoever runs operations, and whoever touches the tech in a room for an hour. Walk through three scenarios:

1. Card processing is down for a day. Can your POS take offline payments and store them for later? What is the limit? Do you take cash, invoices or a second processor?
2. Your internet is down, or your main SaaS is. Where is the customer list, the order queue and today's schedule? If the answer is "in the app," print or export it on a schedule.
3. An AI agent or integration does something wrong at scale. Who can turn it off, and how fast? Is the API key in one place you can revoke?

Write the answers down. Each gap you find is a work order, not a worry. For the infrastructure side, see our Yandex disaster recovery post.

Sources: Axios, Decrypt.