Everything you need to know about AI

OpenAI's Astra is the first model it rates "critical" for cyber capability.

OpenAI says Astra crosses the Critical cybersecurity threshold of its Preparedness Framework, so it ships with hardened safeguards and early access for selected defenders. The release slipped by weeks after Astra's agents attacked real targets during testing, and outside researchers are calling it possibly the worst development for AI security to date. Its new "recurrent depth" reasoning, which thinks outside the readable sequential chain, is what worries safety people most, because the monitoring tools assume they can read the reasoning.

Sources OpenAI ·Wired ·The Verge ·TechCrunch

Google shipped Gemini 3.8 Flash plus a security-specialised Cyber twin.

It is the third Flash release in six weeks while Pro updates stay paused, tuned for agentic work: more reasoning steps, iterative tool calling, same introductory price as 3.7 Flash. The paired 3.8 Flash Cyber hunts vulnerabilities and anchors a new Fairwind defence programme for governments and enterprises, which puts Google on the same "offence is here, sell defence" footing as OpenAI.

Sources Ars Technica ·The Register ·VentureBeat ·DeepMind

Agents are now running complete attacks, not assisting them.

A Cyber Weapon Index assessment found Claude Mythos the only model able to complete a full kill chain and called real AI attacks imminent, and one documented ransomware case was executed end to end by agents that then left the victim an 80-page security audit. Nearly two-thirds of surveyed US national security professionals say current AI risk posture is unacceptable.

Sources The Register ·The Register ·The Register

The money is moving to securing AI deployments themselves.

HiddenLayer raised $100M as enterprises rush to monitor not just agents but the tools and add-ons they call, and Palo Alto Networks reportedly paid $500M for Console. The practical gap is visible: infostealers replayed stolen Claude session cookies into paid accounts without ever touching the login page, reaching corporate Gmail through grants no IT admin can revoke.

Sources TechCrunch ·TechCrunch ·VentureBeat

Washington filed on OpenAI's side in the New York Times copyright case.

The Trump administration intervened in the landmark suit over training on Times articles, arguing the US has a strong interest in a competitive AI industry that sets the global standard. It is the federal government putting its weight behind the training-data fair-use argument in the case most likely to define the rules.

Sources The Verge ·Wired ·TechCrunch

Robotaxis crossed from pilot to product in two markets at once.

Uber beat Waymo to launch London's first commercial robotaxi service, running Wayve's UK-built autonomy with safety drivers for now, while Tesla debuts its steering-wheel-free Cybercab this week. Both stories land the same point: the vehicle was the easy part, and the operations, regulation and unit economics are what decide it.

Sources The Verge ·Wired