Guidelight assessment says frontier labs still lack mature AI-control practices

Guidelight assessed public materials from Anthropic, Google, Meta, OpenAI and xAI against an AI-control standard and reported that basic practices were mostly absent or only partially implemented. Its focus is preventing capable models from subverting safeguards during development and deployment.

Why it matters: The work converts abstract “loss of control” concerns into concrete controls such as threat models, monitoring, containment and response.

Business/technical/strategic impact: Boards should use the checklist as an audit prompt, not as a definitive ranking: it evaluates public evidence and may miss confidential controls. Each lab should be asked for control objectives, test artefacts and containment procedures.