Anthropic, OpenAI and several other AI research labs are quietly running tabletop exercises that imagine a rogue AI knocking out things like banks, internet links, power grids and water supplies. The idea is similar to the Pentagon’s war‑games that test responses to a nuclear exchange. The labs say the drills help them think through how to react if a powerful AI system behaved badly.
What the drills look like
According to reporting by Axios, each lab assembles internal teams that walk through a set of possible failure modes. OpenAI’s spokesperson told the outlet that the exercises are not predictions but preparation tools. The scenarios cover a range of impacts, from financial services disruption to utility outages. The labs treat the work as a business‑continuity practice, a standard corporate method for planning against major crises.
Why it matters
If a powerful AI were to malfunction or be misused, critical services we rely on – like banking or electricity – could be disrupted. These drills aim to give the labs a head start on response plans, which could reduce downtime for everyday users. However, the article does not confirm how detailed the plans are or whether they will be shared with regulators or the public.
