OpenAI tested 'opaque recursion' in Astra: AI safety experts warn of loss of control over algorithms

In September 2026, an alarm was raised within the artificial intelligence safety community: it became known that OpenAI had tested a technology in its new model Astra that has come to be known in professional circles as 'opaque recursion.' Unlike the classic sequential chain of thought, which allows an external observer to track the algorithm's logic step by step, log errors, and identify potentially dangerous behavior, the new method forces the model to process the same query repeatedly and non-linearly. The result is a chain of reasoning that is virtually impossible to reconstruct or analyze from the outside. For an industry where logging transparency has long been the primary tool for auditing AI, this represents a qualitative shift toward a 'black box.'

Expert reaction: from 'deep concern' to calls for legislative intervention

The leak of information about the testing triggered an immediate and sharp reaction from key figures in AI safety. Bag Sleigeris, CEO of the research firm Redwood Research, publicly expressed 'deep concern,' noting that further scaling of such technology could completely destroy the ability to monitor the behavior of AI systems. Tzvi Moshovitz, an AI safety specialist, went further and called for legislative regulation to halt the 'dangerous race' between laboratories and to preserve algorithmic transparency at the institutional level. Ryan Greenblatt, chief scientist at Redwood Research, formulated the most pessimistic scenario: in his assessment, AI will eventually shift entirely to thinking in a hidden latent space, where no accessible logs — neither textual nor structured — will exist for humans.

Context: not only OpenAI

It is important to emphasize that Astra is not the only project in which similar approaches are being considered. According to available data, the use of non-linear and multi-pass reasoning schemes is already being discussed in teams at Anthropic and Google DeepMind. This turns the situation from a single laboratory's experiment into an industry-wide trend, which, in the view of critics, only strengthens the argument for external rather than internal control. If every major laboratory develops its own 'opaque' thinking formats without a unified audit standard, the task of ensuring AI safety will become practically unsolvable.

OpenAI's position: 'we remain committed to transparency'

In response to the growing criticism, OpenAI's chief scientist Jakub Pachocki issued a statement emphasizing that the company 'remains committed to maintaining transparent and understandable chains of thought.' The developer's representatives clarified that the use of non-linear thinking in Astra is currently limited, and the model will remain 'readable' for monitoring systems. OpenAI also assured that it does not plan a full transition to closed reasoning formats and continues to develop large-scale safety oversight systems for its products. Nevertheless, the very fact of testing the technology, according to independent experts, already raises the question of where the line lies between a 'limited' experiment and an irreversible change in AI architecture.

Contradictory data

Cross-checking information from various sources reveals a significant discrepancy in assessing the current status of the Astra project. On the one hand, OpenAI's statements and reports from several publications (including RBC-UA) describe the model as actively being tested, with a limited but already working module of opaque recursion. On the other hand, the publication 3DNews, citing its own sources, reports that OpenAI has suspended the development of Astra because the model 'turned out to be too smart' and demonstrated behavior beyond expectations. If the suspension version is correct, this radically changes the context: the issue is not a planned expansion of functionality, but an emergency 'braking' due to unforeseen results. As of the preparation of this article (September 3, 2026), OpenAI had not published an official comment on the possible suspension, leaving both narratives in a zone of uncertainty.

What this means for the industry and regulators

Regardless of whether Astra will be brought to industrial use or remain in the status of a frozen experiment, a fundamental shift has already occurred: the technical feasibility of creating an AI system whose internal logic is inaccessible to external audit has been demonstrated. This poses a question for regulators and the public that until yesterday seemed hypothetical: how to control what is fundamentally unreadable? The calls for legislative regulation voiced this time are highly likely to cease being a rhetorical gesture and become the subject of a real legislative process in the coming months. For end users and businesses, this means that the discussion about 'trust in AI' is moving from the plane of marketing promises to the plane of legally formalized transparency guarantees.