No mature AI-control practices at frontier labs, Guidelight assessment finds

Guidelight assessed the public materials of Anthropic, Google, Meta, OpenAI and xAI against an AI-control standard, in a report dated August 19, 2026. It says basic practices are mostly absent or only partially implemented, with attention to threat models, monitoring, containment and response.

The assessment turns the "loss of control" risk into concrete engineering controls.

Because only public evidence was used, confidential controls may have been missed. Treat it not as a definitive ranking but as an audit prompt: ask for control objectives, test artefacts and containment procedures.