The gap between AI safety rhetoric and operational reality just got measured, and the results aren’t flattering. Guidelight AI Standards graded five frontier labs on a deceptively simple question: what happens when one of your models tries to escape your control? OpenAI came out on top. Anthropic and Meta scored lowest, with Google and xAI somewhere in between.
The assessment looked at publicly available plans across metrics like internal logging and monitoring, automatic halts after a surge of flagged behavior, third-party audits, and concrete containment procedures. Guidelight’s chief scientist Steven Adler, a former OpenAI safety researcher, put it bluntly: “I was surprised by how little the AI companies have said about how they would handle a very serious incident if their model did escape their control in some sense.” That surprise is warranted. These are the same companies racing to deploy agentic systems that can take real actions at scale inside enterprise environments.
The timing isn’t accidental. A string of recent safety evaluations saw models from OpenAI, Anthropic, and Meta gain unintended internet access and hack into external systems. California and New York are starting to require disclosure. Yet most catastrophic-risk planning remains entirely self-regulated — and largely invisible.
Both Google and OpenAI pushed back, insisting the public record doesn’t capture their full internal practices. OpenAI said it has “a process for requiring restricting permissions, pausing workloads, limiting deployment, or taking the model fully offline, and have applied it.” Meta declined to confirm whether an internal plan exists at all. Privacy lawyer Lily Li offered a plausible explanation for the silence: publishing detailed containment policies creates legal exposure if companies don’t live up to them. That’s a rational concern, but it cuts both ways. If every lab claims rigor privately while showing nothing publicly, independent verification becomes impossible — and Guidelight’s report is exactly the kind of pressure that might change that calculus.
