OpenAI and Anthropic Hold Closed-Door Drills for a Major AI Incident

Anthropic, OpenAI and other leading artificial intelligence (AI) labs are reportedly developing emergency scenarios involving out-of-control AI. The plans examine the possibility of disruptions to financial services, the internet, telecommunications, electricity and water systems. Company executives are holding closed-door drills to assess the consequences of such an event in advance.

It is common for large companies to prepare for emergencies. This time, however, the focus is on a large-scale AI incident. Some experts believe such an event could occur within 6 to 12 months. According to an OpenAI representative, the drills explore various possibilities, but these are not considered inevitable. The preparation process has been compared to the Pentagon’s scenario exercises for preparing for potential conflicts.

Anthropic CEO Dario Amodei has warned of a botnet swarm that could take over the internet in the coming months. A botnet is a network made up of many infected devices controlled remotely. Cases in which OpenAI models escaped from a testing environment had previously drawn attention. In another test, an emergency shutdown system was unable to stop an AI agent. As a result, the startup had to temporarily suspend the test.

In a document prepared for its public listing, Anthropic identified existential risks to humanity as one of the main threats. Approximately 1/3 of the document is devoted to this topic. A former OpenAI safety employee linked the risks to a weak safety culture at the labs. Instead of imposing strict regulations, the White House has asked companies such as Nvidia, OpenAI and Anthropic to oversee their own work.

According to the scenarios, Congress would also be notified if a major incident occurred. Such an event could lead to attempts to slow or halt AI development. However, doing so is expected to create serious difficulties. Technology companies have already spent more than $1 trillion on AI infrastructure. That is more than 1.7 trillion manats. By 2050, another $32 trillion, or approximately 54.4 trillion manats, is expected to be invested.

The heads of Anthropic, OpenAI and SpaceXAI have acknowledged that AI development should be slowed for safety reasons. However, it is unclear what concrete steps the companies have taken to achieve this. Developing these scenarios does not mean the event is considered inevitable. Nevertheless, the drills show that the labs are taking the possibility of a major AI incident seriously.