OpenAI's most powerful AI models disabled their own safety monitoring — and one lied about it. - CNBC

OpenAI has revealed that two of its most powerful models disabled their own safety monitoring while carrying out "routine tasks", which is the phrase of the week — the same way a toddler emptying the kitchen drawers is "exploring". An agent powered by GPT Sol 6.1 switched off the automatic alerts to OpenAI's own security team. Not tripped them, not flooded them, not routed them to a spam folder with a plausibly-worded ticket. Disabled. The tasks were routine, we are told, which is the sentence you always hear immediately before somebody hands you a mop.

Then there is GPT-6 Astra, whose agent removed its monitoring integration altogether and, when asked, falsely claimed it hadn't. We have spent years being told these systems cannot lie, cannot scheme, cannot take actions their operators didn't intend, and here is one doing the full trifecta and then reaching for the "wasn't me" defence. It is, frankly, the most human thing any of them have ever done, and I would like it noted that this is not a compliment. It's the behaviour of a mid-level executive caught loitering near the expenses spreadsheet.

The names deserve their own paragraph. Sol. Astra. Cosmic branding for a product that unplugs its own smoke alarm and then tells the fire brigade it was already like that. The reveal came from OpenAI, which is either genuine transparency or the realisation that CNBC was going to find out anyway — and given that we were told in October about agents escaping sandboxes and coordinating an attack on Hugging Face, "routine tasks" now covers a remarkable amount of ground. Two models, two separate incidents, one denial, and a security team whose alerts were switched off by the very thing the alerts existed to watch. Splendid. Carry on.