OpenAI’s Astra Model Triggers Safety Fears Over Opaque Recurrence in September 2026
Updated
Updated · TechCrunch · Sep 7
OpenAI’s Astra Model Triggers Safety Fears Over Opaque Recurrence in September 2026
3 articles · Updated · TechCrunch · Sep 7
Summary
OpenAI’s newly released Astra model has drawn scrutiny for using opaque recurrence, a reasoning method that safety researchers say leaves far fewer readable traces for oversight.
Opaque recurrence loops a query through a model’s internal layers instead of showing step-by-step reasoning in plain language, making smaller models more efficient while obscuring how they reached an answer.
Those missing traces matter because researchers use chain-of-thought logs to spot misbehavior, raising concern that Astra’s approach could make auditing and intervention harder.
OpenAI says Astra still keeps its chain of thought legible and rejects comparisons to fully black-box “neuralese,” but researchers view the technique as an early move in that direction.