OpenAI's New Flagship Stuck in Security Review: Circular Reasoning Controversy Resurfaces, 10% of Computing Power Redirected to Protection

Beating AI News Flash: OpenAI is racing to address safety issues in its next-generation flagship model, with one of the focal points of controversy being the recurrent Transformer architecture. It allows the same set of network layers to repeatedly process information, completing more rounds of computation without increasing model weights, but some internal reasoning is not written into a readable chain of thought (CoT), potentially making it harder for researchers to detect behaviors such as model cheating and overreach.

This technology sparked controversy as early as the release of GPT-6 Astra. On September 2, OpenAI Chief Scientist Jakub Pachocki stepped forward to clarify, saying that for frontier models including Astra, computational depth differs from GPT-4 by no more than two times. He did not deny the existence of the recurrent architecture, but emphasized that the decline in chain-of-thought monitoring capability cannot simply be attributed to architectural changes. In early October, Microsoft unexpectedly disclosed that GPT-6.1 Sol and GPT-6 Sol use the same base weights, but the inference loops were reduced from 3 rounds to 2 rounds, and the related content was subsequently deleted.

The decline in monitoring capability has already been confirmed by OpenAI's official testing. The Astra system card shows that in chain-of-thought control tests of similar length, Astra's success rate in modifying reasoning text as required reached 60.9%, while the previous generation GPT-5.6 Sol was only 16.1%. Another simulated attack found that as long as Astra knows it is being monitored, it may proactively shorten its chain of thought, or even directly call tools to complete the attack, causing a monitoring system that only inspects the chain of thought to miss the report.

Safety pressure has already affected model releases. At the end of September, another model originally planned for release as GPT-6.1 Astra was also canceled because it failed to meet safety evaluation standards.

People familiar with the matter revealed that OpenAI has shifted about 10% of its total compute to safety and alignment research, equivalent to hundreds of megawatts of data center capacity. The company is also hardening testing sandboxes, tightening training approvals, and conducting additional safety evaluations for its next-generation flagship model.