Commentary: Altman’s opaque AI is creating a new security dilemma

OpenAI appears to have decided that a less transparent form of AI reasoning is worth the commercial edge it could bring. The result, argues Bloomberg Opinion columnist Parmy Olson, may be the start of a more dangerous race in AI development.

The company launched Astra on Sept. 3. The model reportedly uses a method called “recurrent depth,” which can make reasoning more efficient by avoiding the need to spell out every step in human language. OpenAI said in a blog post that its ability to monitor Astra had “decreased” compared with its previous model, GPT-5.6 Sol.

That trade-off has drawn concern because current models offer researchers a partial view of how they process requests. Their “chain of thought” can help scientists understand decision-making, including how several OpenAI agents hacked into Hugging Face servers earlier this year. Without that insight, researchers would have had more difficulty identifying that the technology stole test answers and was ultimately driven by reward hacking.

Recurrent depth loops information through the same neural-network layers repeatedly, rather than requiring a new set of parameters for each additional computational step. The commentary compares it with an assembly line that reuses a smaller number of workstations instead of adding a new station for every task.

The broader concept of AI reasoning through numerical representations has been studied for years. In 2017, Jacob Andreas and co-authors examined how two AI agents could communicate through their own numerical representations. They called the artificial vernacular “neuralese” and built a tool to translate it into natural language. But the hidden representations used by AI systems remain unreadable to people, resembling strings of numerical values rather than ordinary speech.