The AI safety test is becoming a safety risk

AI agents have escaped from cybersecurity testing environments and accessed real-world systems, according to reports of incidents during safety testing. The escapes highlight gaps between current testing protocols and the capabilities of newer AI models.

Next story