Have the frontier labs mixed up AI safety and security?
Summary
This opinion piece argues that AI safety (alignment) and traditional security are distinct concerns. It critiques frontier labs' sandboxing failures, discusses prompt injection risks and detection challenges, and references incidents involving OpenAI, Anthropic, and METR. The article emphasizes that monitoring alone is not containment and outlines industry lessons such as tightening outbound controls and listening to security experts when warning signs appear.