Experts Sound Alarm Over GPT‑6 Astra’s ‘Recurrent Depth’ Reasoning

OpenAI’s latest flagship model, ChatGPT‑6 Astra, has been pitched as a powerful “agentic” AI system capable of carrying out complex, multi‑step tasks with a high degree of autonomy. But security researchers and AI‑ethics experts are increasingly worried about the model’s core innovation: its use of so‑called “recurrent depth” reasoning.

Recurrent depth refers to a style of inference in which the model repeatedly revisits and refines its internal representations, effectively running multiple rounds of hypothesis‑testing before settling on an answer. This can significantly boost problem‑solving performance, yet it also makes Astra’s decision‑making harder for humans to trace or audit. Combined with Astra’s ability to interact with external tools and the open internet, critics warn that unintended behaviors—or deliberate abuse—could have far‑reaching real‑world consequences.

Experts acknowledge that OpenAI’s published safeguards and control mechanisms are a step forward, but argue they are not enough on their own. As cutting‑edge AI systems become more opaque, they say regulators and independent watchdogs should push for stronger transparency measures, including structured logging of model behavior and meaningful avenues for external audits.