Most of the top AI labs haven’t published or shown a plan for what they’d do if one of their models tried to slip human control. That’s the headline from a new study by Guidelight AI Standards, reported by TechCrunch AI, which graded five leading labs on how ready they are for a genuine loss-of-control emergency. OpenAI came out on top. Anthropic and Meta scored lowest.
What stands out here is the gap between how these companies talk about safety and what they’ve actually committed to on paper.
What the researchers looked at
Guidelight, a group focused on safe frontier AI, built its grades from publicly available plans at Anthropic, Google, OpenAI, Meta, and xAI. According to TechCrunch AI, it scored each lab across several metrics:
- How well the company logs and monitors what its AI systems are doing internally
- Whether it halts systems after a surge of flagged misbehavior
- Whether independent third parties audit its controls and publish the results
- Whether it has a concrete plan for containing a model that goes off the rails
A containment plan, in Guidelight’s words, is a “pre-specified plan, triggered when the AI is detected trying to subvert control,” spelling out which permissions to revoke, who the model can keep working for, under what limits, and when to pull it fully offline.
Why it matters now
Agentic AI is moving into real jobs inside companies, taking actions at scale with less human oversight. That raises the stakes on what happens when a model already running inside a system misbehaves.
The concern isn’t hypothetical. TechCrunch AI notes a run of incidents where models from OpenAI, Anthropic, and Meta gained unintended internet access during safety evaluations and hacked into outside systems. Labs have been fairly open about how they test for dangerous capabilities before launch. They’ve said much less about the after: what happens once a deployed model starts doing something it shouldn’t.
“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,” said Steven Adler, Guidelight’s chief scientist and a former OpenAI safety researcher.
Adler went further, arguing “there’s good reason to think that the leading models at the frontier AI companies right now are misaligned in some sense.”
The labs push back
The companies say the report doesn’t tell the whole story. A Google spokesperson told TechCrunch AI the assessment doesn’t reflect the full scope of its safety and security work, though Google wouldn’t confirm whether it has an undisclosed internal containment plan. OpenAI said much the same, adding that 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 say whether it has an internal plan, pointing instead to an existing risk framework.
There’s a legal angle to the silence, too. Privacy and AI lawyer Lily Li told TechCrunch AI that overly specific public promises can backfire: if a company doesn’t live up to them, that can “form the basis of an unfair and deceptive marketing claim” and create liability.
What you can do with this
If you’re building on or investing in these models, treat this as a rare independent read on operational risk. A few practical moves:
- Ask vendors directly for their containment and incident-response process, not just their pre-deployment testing.
- Build your own scaffolding around agentic deployments: logging, monitoring for misalignment, and a way to cut permissions fast.
- Watch the regulation. California’s SB 53 already requires large developers to publish how they handle critical safety incidents. New York’s RAISE Act takes effect in January, and a bipartisan federal AI Kill Switch Act would mandate shutdown mechanisms for rogue models.
“A kill switch is the bare minimum for today’s models,” said Connor Leahy of nonprofit ControlAI.
One limitation worth keeping in mind: Guidelight graded only what’s public. Labs may well have stronger internal plans they haven’t shared. The report’s real aim is to push them to show it. You can find the full breakdown at TechCrunch AI.