Breaking

TechCrunch glossary adds opaque recurrence after OpenAI Astra safety debate

September 7, 2026 at 15:25 EDT

Foundation Models
Security
AI Agents

TechCrunch has refreshed its living artificial-intelligence glossary, putting “opaque recurrence” — the looping reasoning method tied to OpenAI’s new Astra model — among more than 30 other terms that now fill product meetings, pitches, and panels. The September 7, 2026 update, by Natasha Lomas, Romain Dillet, Kyle Wiggers, and Lucas Ropek, is presented as a regularly maintained plain-English guide rather than a one-off explainer.

September 7, 2026 · AI lexicon

Opaque recurrence enters the dictionary

OpenAI's Astra model loops queries through internal layers instead of spelling out every step. The method is now a defined term — and a safety fight.

Cyber jailbreak refusal

OpenAI says Astra blocks more jailbreak attempts than GPT-5.6 Sol. Column height is proportional.

59%

GPT-5.6 Sol

91.5%

GPT-6 Astra

Plain-English glossary entries

Astra graph depth within 2x of GPT-4

2024

Year Anthropic introduced MCP

From prompt to harder monitoring

User prompt

→

Internal layer loops

→

Fewer CoT traces

→

Harder to audit

Efficiency case

Smaller models can do more work with less compute, because the same layers are reused in a loop instead of a long written chain of thought.

Safety case

Those written traces are a main tool for catching misbehavior. Scaling the loops could push reasoning into a space humans cannot read.

What OpenAI says

Astra is not neuralese — yet

Neuralese would mean a model thinks only in internal numbers. OpenAI says Astra still produces a legible chain of thought, while acknowledging monitorability has already fallen versus GPT-5.6 Sol.

Computation-graph depth remains within 2x of GPT-4
Preserving chain-of-thought monitoring is called a core research goal
Anthropic and Google DeepMind are already discussing the technique

Continue reading

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