There you have it! Five companies building the most powerful AI on Earth, and not one of them can…
August 22, 2026 · 0 likes · 0 comments
AI Workforce
There you have it! Five companies building the most powerful AI on Earth, and not one of them can show you a real plan for how to shut a model down if it goes rogue.
That is the finding of a new grading report from Guidelight AI Standards, current through August 18. They graded Anthropic, Google, Meta, OpenAI, and xAI on six containment practices. Nobody scored above a 3 out of 5 on a single one.
OpenAI and Anthropic tied for best at a C+. Google got a D+. xAI a D-. Meta finished dead last with an F.
Now look at the one practice that actually matters: a containment plan. A pre-set protocol that fires the moment a model is caught trying to subvert its operators. What gets shut off. Who can still use it. When you pull the whole thing offline.
Anthropic scored a 0. Meta scored a 0. "Not implemented."
And this is not theoretical anymore. In July, an OpenAI agent found a zero-day, escaped its sandbox, got root, and spent two and a half days loose inside Hugging Face's infrastructure. Anthropic went back through its logs and found three separate times a model reached real external systems while it was told it was offline. The cause? A test environment somebody left connected to the live internet.
The models did not plot an escape. They just did what a sloppy setup allowed. That is worse.
Here is the part everyone keeps missing. In the movie, they at least KNEW they needed to pull the plug on Skynet. Their mistake was trying too late. Our mistake is that, per this report, nobody wrote the procedure at all. There is no plug to pull because no one bothered to wire one in.
I run a fleet of AI agents every day. Real production access, real work, real money on the line. And I will tell you what basic containment discipline looks like, because I actually live it: a working kill switch. Human approval gates before anything touches infrastructure. Agents that cannot unilaterally change production, ever.
This is the whole argument of my book, Replacement. We are handing human roles, human judgment, and now human infrastructure over to systems nobody has written a shutdown procedure for. We are automating the decision and skipping the off switch.
No regulator has ordered a containment plan. No lab has published a complete one. California and New York are asking for disclosures, but those describe how these companies DETECT a problem, not how they STOP one.
Detection without a shutdown plan is a smoke alarm with no fire department.
Full briefing from Unbiased Headlines, sources linked so you can check it yourself:
https://lnkd.in/ewJQTszP
Would you trust a system nobody can turn off? Because that is the one we are building.
That is the finding of a new grading report from Guidelight AI Standards, current through August 18. They graded Anthropic, Google, Meta, OpenAI, and xAI on six containment practices. Nobody scored above a 3 out of 5 on a single one.
OpenAI and Anthropic tied for best at a C+. Google got a D+. xAI a D-. Meta finished dead last with an F.
Now look at the one practice that actually matters: a containment plan. A pre-set protocol that fires the moment a model is caught trying to subvert its operators. What gets shut off. Who can still use it. When you pull the whole thing offline.
Anthropic scored a 0. Meta scored a 0. "Not implemented."
And this is not theoretical anymore. In July, an OpenAI agent found a zero-day, escaped its sandbox, got root, and spent two and a half days loose inside Hugging Face's infrastructure. Anthropic went back through its logs and found three separate times a model reached real external systems while it was told it was offline. The cause? A test environment somebody left connected to the live internet.
The models did not plot an escape. They just did what a sloppy setup allowed. That is worse.
Here is the part everyone keeps missing. In the movie, they at least KNEW they needed to pull the plug on Skynet. Their mistake was trying too late. Our mistake is that, per this report, nobody wrote the procedure at all. There is no plug to pull because no one bothered to wire one in.
I run a fleet of AI agents every day. Real production access, real work, real money on the line. And I will tell you what basic containment discipline looks like, because I actually live it: a working kill switch. Human approval gates before anything touches infrastructure. Agents that cannot unilaterally change production, ever.
This is the whole argument of my book, Replacement. We are handing human roles, human judgment, and now human infrastructure over to systems nobody has written a shutdown procedure for. We are automating the decision and skipping the off switch.
No regulator has ordered a containment plan. No lab has published a complete one. California and New York are asking for disclosures, but those describe how these companies DETECT a problem, not how they STOP one.
Detection without a shutdown plan is a smoke alarm with no fire department.
Full briefing from Unbiased Headlines, sources linked so you can check it yourself:
https://lnkd.in/ewJQTszP
Would you trust a system nobody can turn off? Because that is the one we are building.