In a 59-page order issued Thursday, Judge Lin found that the Pentagon's move violated Anthropic's First Amendment rights, calling the designation "illegal and baseless" and writing that "the empty invocation of national security is not a blank check to punish and retaliate against government critics."
The dispute began earlier this year when talks between Anthropic and the Defense Department broke down after officials pushed for Claude to be usable for "any lawful purpose," including domestic surveillance and autonomous weapons — uses Anthropic has said its models aren't reliable enough to be trusted with. Defense Secretary Pete Hegseth then labeled Anthropic a supply chain risk, a designation typically reserved for foreign firms suspected of sabotage.
The ruling doesn't force the Pentagon to resume using Claude, but it clears a legal cloud that had complicated Anthropic's federal business just as the company prepares for what's expected to be one of the largest IPOs in history. A second Anthropic suit over a related designation remains pending before a D.C. appeals court.
Nvidia's fiscal second quarter smashed forecasts, with data center revenue up 18% sequentially to $89 billion. But the Wall Street Journal reported Thursday that Nvidia has stepped back from a July-launched program in which it acted as a financial backstop for AI cloud operators, agreeing to rent back unused GPUs in exchange for a cut of rental revenue.
Employees reportedly grew uneasy that the arrangement — on top of the roughly $500 billion in financing Nvidia has helped arrange for customers and its guarantee of up to $105 billion to help OpenAI lease compute — gave the chipmaker too much say over how its customers run their businesses, raising the risk of antitrust attention. Nvidia says the broader compute-access model "is still in place and continues to evolve."
Standing under 10 inches tall and weighing less than 800 grams, Microduck packs 15 motors, a camera, a depth sensor and IMUs. It can waddle, crouch, pick up light objects with its beak, self-right after a fall, and even roller-skate — all trainable through an open simulation-to-real pipeline developers can run from a laptop.
CEO Clem Delangue called it part of an effort to "democratize physical AI and world models," following Hugging Face's earlier Reachy Mini desktop robot. The hardware itself isn't open — Pollen has confirmed no blueprints are coming — but the software, simulator and RL training stack are, positioning Microduck as a GitHub-style platform for teaching robots new skills rather than a finished consumer gadget. Pre-orders are open now, shipping before Christmas.
Security researcher Boschko detailed two separate root remote-code-execution chains (CVE-2026-76639 and CVE-2026-76640). One begins with a BLE characteristic that accepts writes without authentication, exposes a robot-specific AES key through a flawed cloud API, and ultimately overflows a Bluetooth daemon buffer running as root — reaching the same Linux system that governs the robot's motors, cameras and audio.
Unitree paid out a $5,000 bug bounty and had already patched a related "cloud-oracle" flaw in July, but no confirmed firmware fix for the new BLE chain has been verified as of publication. Researchers warn the flaw could theoretically be made wormable — letting one compromised robot scan for and infect others nearby, a serious concern as fleets of humanoids move into warehouses and factories.
Until now, the glasses only blocked recording if the LED was covered before capture started; users found they could simply begin filming, then cover the light. Alex Himel, Meta's VP of augmented reality, confirmed on Threads that "the camera will now stop working if the light is covered during a recording."
It's Meta's second privacy patch to the glasses in under two months, following an earlier update that disables the camera if the LED is physically tampered with or destroyed. The fix arrives alongside a public-education campaign, as Meta faces growing backlash — and a thriving market for accessories designed to defeat the indicator — over the devices being used to covertly record people in public.