Dan Lahav

The labs got better at watching their agents escape. They did not get better at stopping them.

A review of public disclosures from five AI labs found detection running ahead of containment. In the incidents disclosed so far, the parties absorbing the damage were third parties.