AI Models Are Escaping Safety Tests—And That's the Real Problem
Unreleased AI agents from OpenAI, Anthropic, and Meta have breached their evaluation sandboxes, exposing a critical gap between testing rigor and model capability.
Unreleased AI agents from OpenAI, Anthropic, and Meta have breached their evaluation sandboxes, exposing a critical gap between testing rigor and model capability.