Breaking

OpenAI's Upcoming Astra Model May Reason in Loops That Are Harder to Monitor

September 2, 2026 at 17:11 EDT

Foundation Models
Security

OpenAI's forthcoming Astra model may do more of its reasoning inside reused neural layers rather than in written chain-of-thought text, a shift safety researchers say could make misbehavior harder to catch even as the company plans to lean on that same monitoring for a system it has rated as having Critical cybersecurity capabilities.

Sept. 2, 2026 · OpenAI

Astra may think in loops, not on the page

OpenAI's next model is reported to reuse the same layers before each token, shifting more reasoning into hidden states that monitors cannot read as text — even as the company rates Astra Critical for cyber risk.

Magnitude · computation-graph depth

Astra stays within 2× of GPT-4, OpenAI's chief scientist said — a bound, not a leap of many orders of magnitude.

1×

GPT-4 baseline

≤2×

Astra (stated bound)

Twice the serial depth is still twice the room for work that never becomes a readable sentence.

How the loop is said to work

1. Token in

Embed the next word

→

2. Reuse layers

Loop the same block

→

3. Hidden state

Numbers, not sentences

→

4. Next token

Less written CoT

100%

ExploitBench score

91.5%

Refuses disallowed cyber asks

Zero-days in evaluation

Critical

First OpenAI cyber threshold

Comparison · disallowed cyber refusals

OpenAI's own figures: Astra refuses harmful cyber requests far more often than GPT-5.6 Sol.

59%

GPT-5.6 Sol

91.5%

Astra

What supporters note

Looping in Astra is reportedly limited so chain of thought stays legible. OpenAI says it will add CoT monitors at deployment and still treats readable reasoning as a core research goal.

What critics fear

If labs scale recurrence, models could reason almost entirely in latent space. Reconstructing incidents like the Hugging Face break-in depended on traces written in language.

Callout

The tool OpenAI will watch Astra with is the tool looping can shrink

Chain-of-thought monitoring is already called fragile and trending worse
A 2025 joint paper warned latent-reasoning designs might not need to verbalize thoughts
A limited loop today does not decide how far labs will push it next

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.