AI Trend Research
EN

← October 14, 2026 briefing · added in the 20:00 KST update

P2 Model Unconfirmed Language modelsSafety & alignment20:00 updateOpenAI

OpenAI's next model uses a 'loop transformer' design whose reasoning is hard to monitor, putting the focus on safety (report)

Announced October 13, 2026

Verified (fact)

The Information reported on the 13th (local time), citing people familiar with the matter, that OpenAI is focusing on alignment and security work ahead of releasing its unannounced next-generation flagship model. The new model's 'loop transformer' design could make the model's reasoning process harder to monitor, so work is under way to make both the model and its testing environments safe. No release date has been set.

Why it matters (interpretation)

An architecture that passes data through the same layers repeatedly moves reasoning out of CoT (chain-of-thought) text and into internal representations. That could weaken current safety methods that rely on CoT monitoring. This is the first concrete sign of a conflict between next-generation model design and how safety oversight is done.

Confidence medium · official source pending

Sources

More about OpenAI