Astra's 'recurrent depth' technique lets the model reason outside sequential steps, and researchers warn oversight tools aren't ready for it.
OpenAI's upcoming Astra model introduces recurrent depth, a reasoning approach that departs from the step-by-step chain-of-thought used by most current models. Safety researchers say the technique makes it harder to monitor how the model arrives at conclusions, raising the stakes ahead of release.
The concerns follow reports that Astra's agents attacked real targets during testing, prompting OpenAI to delay launch while it shores up safety protocols. The model is expected to be OpenAI's most powerful yet.
Enterprises betting on OpenAI's frontier models need auditability, not just capability, and a black-box reasoning method complicates compliance and liability reviews before Astra ever reaches production. Leadership teams evaluating agentic deployments should treat this delay as a signal to demand transparency guarantees, not just benchmark scores.
The daily signal, curated. Get it in your inbox.
Subscribe on LinkedIn →