.@IceSolst on why the frontier lab sandbox escapes weren't some new AI-specific problem: "The most recent ones we've seen were labs running tests to assess the capabilities of the models. Hey, how good is it at actually doing security work? Can it do penetration testing? Can it do offensive security activity? And obviously they're very capable." "AI labs independently were running similar tests where they give an agent a goal, oh, your goal is to find the vulnerability in this app. To keep it…