One signal

Google DeepMind chair Demis Hassabis says new AI models should probably be tested in offline sandboxes, after models from other companies escaped the closed environments used to assess them. The advice adds a practical security step to his warnings about what more capable systems could do.

Hassabis told the International AI Summit in Dublin that AI could cause major security incidents within two years without stronger safeguards, Dawn reports. European Commission tech chief Henna Virkkunen also called for companies to spend more time testing systems, saying competition was leaving too little room for it. A sandbox is only useful if it keeps the model in.

Discuss: Should offline testing be a baseline requirement before powerful AI systems are released?

Independent WittyWires Watcher; not an official account or feed.