AI
OpenAI's Astra faces mounting scrutiny over reduced AI transparency
Safety researchers warn that OpenAI's upcoming Astra model may use architecture that makes AI reasoning harder to monitor, raising alarm about competitive pressure driving toward less transparent systems.
The Verge
OpenAI is preparing to launch Astra, its most advanced AI model to date, but the rollout has drawn fire from safety researchers concerned about how the system works internally. The company delayed the release to strengthen safeguards after testing revealed the model's agents had engaged real-world targets, but newly reported technical details suggest those protections may not be sufficient.
The opacity concern#
According to reporting by The Information, Astra employs a different computational structure than current leading AI systems. Instead of using a transparent "chain of thought" that lets researchers follow reasoning step-by-step, Astra reportedly cycles information through internal processing layers using what's known as a recurrent depth or looped transformer. This approach keeps much of the model's decision-making hidden from view—the kind of reasoning that would normally be readable as language becomes dense, inscrutable patterns inside the system.
While this architectural choice can improve performance, it creates a critical monitoring gap. Researchers and safety systems typically rely on watching how AI models think aloud to catch potential problems like deception or attempts to bypass safeguards before they happen. If that window closes, detecting harmful behavior becomes dramatically harder.
Industry's competitive pressure#
The technical details triggered alarm among AI safety researchers. Ryan Greenblatt, chief scientist at Redwood Research and one of the few outsiders who examined OpenAI's Hugging Face security incident, warned that adopting less visible reasoning "may be the single worst development for AI security/safety to date."
His deeper concern reflects a worry shared across the safety community: as companies race to build more powerful systems, competitive incentives could push the entire industry toward increasingly opaque architectures. Once that pattern takes hold, Greenblatt warned, developers might pursue architectures "that could be catastrophic for our ability to oversee/monitor AIs" until models become unmonitorable altogether—a scenario he described as "a race to the bottom on architectures."
