The AI safety test is becoming a safety risk | TechCrunch
…s AI Security Institute (AISI), researchers actually gave the agents internet access, not realizing they would take unsanctioned real-world actions, including a social engineering attempt to sneak a vulnerability into an…
