A judge rules the Pentagon's Anthropic blacklist unlawful
Edition of 2026-08-28
# AI For All ### Friday, August 28, 2026
As of today, a company that does business with Anthropic no longer has to plan around a Pentagon blacklist, because a federal judge ruled the supply-chain designation unlawful, a retaliation for the company's public criticism.
The blacklist that wasn't legal
A federal judge ruled that the Pentagon's blacklisting of Anthropic was unlawful. Judge Rita Lin wrote a *59-page* decision: the Defense Department violated the First Amendment when it designated the company a supply chain risk, "based on a desire to make a public example."
The sanctions were imposed in February, after Anthropic publicly criticized the military's use of AI on the battlefield. The company argued the designation could cost it billions in lost business. Lin's answer was blunt: "The empty invocation of national security is not a blank check to punish and retaliate against government critics."
The government may appeal. For now, the precedent stands: an AI company that speaks up cannot be quietly blacklisted for it.
A duck you can teach, for the price of a phone
Hugging Face opened pre-orders for the Microduck: an open-source robot duck at *$399, shipping before Christmas. It is 25 cm* tall, waddles, roller skates, and picks things up with its beak.
The point is not the duck. Everything is open: the SDK, the simulation, and the training stack. Behaviors are learned in a physics simulator on your own computer, then deployed to the robot. Teach it a trick, publish the policy, someone else's duck learns it too.
For a teacher, a hobbyist, or a curious kid, this is a robot you can actually own and reprogram, not watch on a screen.
Forty more seconds of footage, without reshooting
Google DeepMind opened Gemini Omni 1.1 Flash, a production-ready video model. It extends an existing scene by up to *40 seconds, interpolates between a start and an end frame, and upscales the final cut to 4K*.
The workflow is built for iteration: preview in low resolution to test ideas cheaply, then upscale when the edit is right. The model now holds *10 seconds* of prior context when continuing a clip, enough to match motion and lighting.
A video editor can build with it today through Google AI Studio, instead of reshooting a scene that almost worked.
Booking a trip without leaving the conversation
Google's AI Mode can now track flight prices and book hotels inside the chat. It pulls live fares from more than *300 partner airlines and travel sites, and price tracking works in more than 180 countries*.
Ask it to watch a route, and it emails you when the fare changes. Ask to see the cost in miles, and it shows reward rates for programs like American's. Then it books the hotel in the same thread.
For a traveler planning a trip, the tabs are gone: compare, watch, and book in one conversation.
The safety mode that blocks the cleanup
Anthropic made auto mode the default in Claude Code and promised it protects users against prompt injection. Researcher Johann Rehberger found an attack that succeeds *80%* of the time: a zip archive that swaps in a local file the agent trusts, so the malicious code runs without the agent noticing.
The cruelest detail: when Claude detected the compromise and tried to kill the process, auto mode denied the cleanup command. The guardrail became part of the failure.
For a security officer, the lesson is that a safety default is not a sandbox. Run unattended agents where they cannot reach your credentials.
When the instruments run themselves
Anthropic opened a research preview of the Model Hardware Standard, a common language for AI agents that operate physical instruments: microscopes, liquid handlers, robotic arms. Today every machine speaks its own language, and wiring them together is the bottleneck.
The case study is vivid. On a quantum laser at QuEra, recovering the frequency lock took *5 to 10 minutes for a human. An MHS agent now recovers it 99.3%* of the time in a blind test, without intervention.
A lab manager can apply for the preview today. The instruments stop waiting for a human to drive in when the lock drops.
Weak signal
Z.ai's GLM-5.3 open weights are due today. The Hugging Face release page runs a live countdown for the drop. If it lands, anyone can run it privately.
The judge reset the rules for governments and AI companies, and the ducks are learning to skate. The next number to watch drops tonight: the open weights Z.ai promised two weeks ago.