Executives at Anthropic, OpenAI, and other AI companies are privately gaming out scenarios for how to respond if a catastrophic AI incident triggers a public and political backlash.

These executives anticipate a large-scale accident, most likely a cyberattack, that disrupts financial services, network communications, and even water and power supplies. Once unsafe AI causes major real-world harm for the first time, an already wary public would turn further against the technology and those steering it, including Anthropic CEO Dario Amodei and OpenAI CEO Sam Altman; criticism would also target President Trump for his reluctance to introduce regulation. Scenario gaming is common in many corporate and national security fields, and the U.S. Department of Defense has conducted war games for decades. What is different this time is that many leading AI researchers and executives believe a major incident is inevitable. An OpenAI spokesperson said: "OpenAI conducts emergency drills, and teams discuss and game out responses to a range of potential scenarios. We do not believe these scenarios will necessarily happen; the purpose of the drills is to help us respond to various unexpected situations." Anthropic declined to comment.

Whether a rogue-agent swarm escapes from an internal testing environment or malicious actors find unexpected ways to exploit existing large models, a blame game would begin immediately. Recent attacks on South Korean financial institutions, with two banks reportedly breached, have already demonstrated the level of damage a single person can cause. After the midterms, a rising Democratic force would quickly move to restrict AI, but it would face enormous resistance. Aging members of Congress lack understanding of AI-driven change and are likely ill-equipped to handle the situation. The global economy is now deeply tied to AI infrastructure buildout. Slamming on the brakes would deal a heavy blow to an already fragile economy. Many open-weight models can be downloaded for free and used for malicious purposes, and most cybersecurity practitioners believe this problem can only be countered by using AI against runaway AI.

This playbook includes red-teaming exercises for extreme worst-case scenarios, with the core work being to quickly educate members of Congress about AI. Corporate executives know that passing relevant regulatory legislation right now is nearly impossible, but they still hope to shape the policy framework in advance so that once the first major disaster erupts, the U.S. political world will use this playbook to legislate. Worth watching: multiple AI industry insiders say they believe a major incident will occur within the next 6 to 12 months. Even if Democrats control both the Senate and the House, companies judge that political division will remain the norm, and divisions within the Democratic Party are also severe. Some hardline Democratic proposals include banning superintelligence and pausing advanced AI development. Other proposals have bipartisan support and are even accepted by the industry to some degree, such as requiring mandatory emergency kill switches for advanced AI. But experts question whether it is practically possible to shut down all AI systems entirely. When a real crisis arrives, Washington can set aside partisan differences. A Democratic staffer cited bipartisan cooperation during the COVID-19 pandemic and the 2008 financial crisis as evidence.

Terminology notes

1. Scoop: a journalism term for an exclusive flash or exclusive tip

2. gaming out scenarios: scenario gaming, simulation playbooks

3. rogue-agent swarm: a swarm of out-of-control agents

4. red-teaming: red-team/blue-team confrontation, security stress testing

5. kill switch: emergency shutdown switch, termination switch

6. open-weight models: open-weight models

7. midterms: U.S. midterm elections

8. Axios: Axios, a U.S. political news website

Note: References to hackers in the original text are one-sided claims by foreign media, and the relevant facts remain to be verified. Would you like me to compress the translation into a news summary of 200 words or fewer?