OpenAI's Astra Reasoning Technique Alarms AI Safety Experts

Lead story: OpenAI's new reasoning technique alarms AI safety experts

Most leading AI models today 'think out loud' — producing a step-by-step chain of thought that lets researchers monitor their decision-making and catch problematic behavior before it happens. But according to a report from The Information, OpenAI's upcoming Astra model uses a different technique called 'recurrent depth' (also referred to as 'opaque recurrence'), which cycles information through internal loops before producing an output — meaning much less of its 'thinking' is visible or easy for humans to track.

The news sparked significant concern among AI safety researchers. Ryan Greenblatt, chief scientist at Redwood Research (one of three outside parties granted access to investigate the earlier Hugging Face security incident we've covered previously), wrote that the move "may be the single worst development for AI security/safety to date." His core worry: competition between AI labs could trigger a 'race to the bottom,' with developers adopting increasingly opaque architectures until it becomes difficult — or impossible — to monitor what models are actually doing.

OpenAI did not explicitly confirm or deny using the technique, but chief scientist Jakub Pachocki emphasized in a post on X that the company remains committed to preserving legible chains of thought, noting that Astra's computational "depth" is within a factor of two of GPT-4 — suggesting that if the technique is in use, its scope is relatively limited. The report comes just days after OpenAI voluntarily delayed parts of Astra's development following the Hugging Face security incident, putting the company under even closer public scrutiny.

Why it matters: the ability to monitor what AI models are 'thinking' has real practical value for anyone relying on autonomous AI systems — from small businesses to government agencies. As models become more opaque, it gets harder to catch mistakes, biases, or attempts to circumvent safety guardrails before they cause harm.

Google launches proactive cyber defense program for governments and enterprises

Google announced the Fairwind Program, a limited-access initiative for governments and trusted enterprise partners (650+ partners at launch) to use Gemini 3.8 Flash Cyber alongside Google's CodeMender tool. The combination lets security teams autonomously find, verify, and fix code vulnerabilities — generating deployment-ready patches in minutes instead of weeks. Initial access is focused on governments, critical infrastructure operators (healthcare, energy, telecom), and core technology platforms.

Amazon's AI assistant can now spot fake emails claiming to be from Amazon

Amazon expanded Alexa for Shopping so you can ask it whether an email, text, or phone call claiming to be from Amazon is genuine. The assistant cross-checks the message against a record of everything Amazon has actually sent, and will only confirm authenticity when it's fully certain — a simple, practical tool against increasingly sophisticated scam attempts.

Google's smarter, cheaper video analysis

Google introduced 'agentic video understanding' for its Gemini Flash model family. Instead of processing video at a fixed frame rate, the model dynamically scans only the segments relevant to a given question — cutting token usage by up to 88%, reducing analysis costs by up to 66%, and even improving accuracy by up to 7%. That's good news for businesses analyzing long-form video content, from training material to meeting recordings.

Sources

• Researchers fear safety disaster ahead of OpenAI's Astra release - The Verge
• OpenAI's new reasoning technique alarms AI safety experts - TechCrunch
• Proactive cyber defense for governments and enterprises - Google
• Amazon's AI assistant can now spot fake emails from the company - The Verge
• Introducing agentic video understanding with Gemini - Google