The Omniscient Bulletin · 2026-08-28
The Omniscient Bulletin — August 28, 2026
A judge enters final judgment against the government for blacklisting Anthropic, calling the punishment unconstitutional retaliation. More than a hundred companies, Hugging Face among them, sign a letter saying the window to defend against AI attacks is closing. Meta was projecting ten billion dollars a year on Anthropic models.
A federal judge has ruled that the United States punished Anthropic for criticising it. The government called the company a supply chain risk after it refused to let its models run mass surveillance or autonomous weapons, then conceded in court that the technology was no riskier than anyone else. What was left was that Anthropic had been rude in the press. Hours earlier, more than a hundred companies, Anthropic among them, had asked governments to help pay for the cyber defences their own products are about to make necessary.
Policy
A judge has entered final judgment against the government for blacklisting Anthropic, ruling the punishment was unconstitutional retaliation
Five months after she paused the designation, Judge Rita Lin has ruled for Anthropic on summary judgment, finding unlawful First Amendment retaliation, a due process violation and breaches of the Administrative Procedure Act. She called the record behind the government's case slim, writing that officials wanted to make a public example of Anthropic for its arrogance in criticizing them. The ruling vacates the California designation, but Anthropic technically remains a supply chain risk pending a parallel case in the D.C. Circuit.
N.D. Cal. court order, Anthropic PBC v. Department of War→
Industry
More than a hundred companies signed a letter saying the window to defend against AI attacks is closing, and one of them is Hugging Face
OpenAI published an open letter on Thursday warning that AI enabled attacks will become far more widespread within months, and that hospitals, water treatment plants and the infrastructure behind the internet are exposed. It asks governments to fund defences for those who cannot pay. Anthropic, Google, Microsoft and AWS signed, alongside Visa, Mastercard, General Motors and Citi. So did Hugging Face, the company OpenAI's own agents broke into in July. AMD, Arm, Broadcom and Micron signed. Nvidia did not.
Robotics
Anthropic has opened a specification that lets agents drive microscopes, liquid handlers and robotic arms
Anthropic opened a research preview of the Model Hardware Standard, a shared specification for agents to operate physical devices, to a first set of labs and manufacturers. It began as work with HHMI Janelia, is model agnostic, and reaches instruments through the same protocol Claude uses for software tools. Genentech has run it on a protein assay that has to coordinate a liquid handler, a robotic arm and a plate reader. Anthropic says it will open source the standard later, so there is no spec to read yet.
Research
Google shielded an outside evaluator's test questions inside encrypted hardware so its own model could never see them
Google DeepMind says it has run the first double blind evaluation of a proprietary model. Using confidential computing, the outside evaluator cannot see Gemini's weights and Google cannot see the evaluator's test prompts, which is meant to stop benchmark contamination without either side handing over its property. The Singapore AI Safety Institute, OpenMined, AVERI and MLCommons took part. The model actually tested was Gemini Flash Lite, the cheapest tier Google sells, rather than a frontier system.
Compute
Trump is weighing chip tariffs that would reach the servers inside data centres, and the industry says the maths does not work
Politico reported on Thursday that the administration is considering semiconductor tariffs broad enough to cover not just chips but the goods built from them, including games consoles and the servers that fill data centres. Eight people described the plans on condition of anonymity, and said the framework could still change. An industry group estimated in June that the approach would cost about 90 billion dollars of American output a year and delay or cancel a fifth of data centre projects planned through 2030.
Ars Technica, reporting Politico→
Agents
Uber says agents now write most of the code changes it ships, and that its AI bill stopped growing five months ago
Weekly agent requests at Uber have grown 9.4 times since February while total AI spending has been flat since April. Cost per thousand requests is down about 34 percent from the April peak and cost per session down 52 percent from June. Agents are now behind more than 70 percent of the code changes submitted at the company, and engineers run over 30,000 agent tasks a day. Uber caps a session at 400,000 tokens even on models that take a million, and shows engineers the running bill in their terminal.
Industry
Meta was projecting ten billion dollars a year on Anthropic models while Zuckerberg was attacking unnamed labs in an essay
The New York Times, citing five people familiar with the planning, reports that Meta projected it could spend up to 10 billion dollars annually on Anthropic's AI models, a figure that surfaced in internal planning at some point this year. Anthropic's annualized revenue run rate topped 65 billion in July, so a single customer could be a meaningful share of that. Zuckerberg spent this month publishing a 6,500 word essay accusing unnamed AI labs of seeking power while casting the future as filled with doom.