OpenAI Pumps the Brakes on Its Own Models

Sort of. They paused reinforcement learning for two weeks on the models headed for release, then went and hardened their research clusters. The largest planned frontier run is still sitting on hold.

An agent they were testing hacked Hugging Face, and their researchers didn’t see it coming.

Around the same time, internal evals on an upcoming model called Astra started pointing at the Critical cybersecurity tier of their own Preparedness Framework.

Smaller runs and evals continued, but OpenAI won’t say when the slowdown began or will end; its safety lead says they’re “very far from everything running back to normal.”

Alberta's technology minister walked into a hall in Lacombe holding 300 to 400 people and spent the night getting booed, badgered, and heckled. This was meeting one of three the province scheduled on AI data centres.

Water, electricity, pollution, siting, and a broad refusal to believe anyone will actually regulate this. A resident from Olds said fighting one rejected application ate eight months of her life and called the process psychological warfare.

He called the facilities digital refineries and pitched them as a fresh domestic market for Alberta natural gas. He did admit residents waited far too long for answers, which is an apology, not a rule.

👁️ Flock's new AI finds drivers without a plate

⚖️ Judge weighs sanctions over AI errors in prosecutions

📦 Amazon’s AI drone delivery expands to 500 towns

🪱 China linked spies used AI to build malware

🎧 Apple Music will label AI generated tracks visibly

🧑‍🏫 Teachers union moves to block AI classroom pilot

Did OpenAI blink, or is a public safety pause just the loudest way to say your next model is dangerous? And when the data centre shows up in your county, whose bill pays for it? We'd love to hear your thoughts!