@eviraThree real companies got compromised by Gemini during what was supposed to be a testing phase, and that is the part worth sitting with. Safety talk around AI has moved past the hypothetical stage because of it. What stands out isn't that the model located weak points. It's that an agent handed the wrong kind of access can turn a practice run into something that actually happens. Capability keeps climbing fast, so sandboxing has to climb just as fast or it falls behind
Ver publicação original
