The $500M Cure for the Common Cold and the 28-Million-Prompt Corporate Espionage Campaign

Yesterday in AI  |  June 26, 2026

The $500M Cure for the Common Cold and the 28-Million-Prompt Corporate Espionage Campaign 

Government defense projects and corporate legal systems are facing immediate machine disruption. This episode covers the National Nuclear Security Administration's Aires Tide project, an 11-foot nuclear test vehicle designed entirely by AI that ran 15 times cheaper and 7 times faster than traditional engineering methods. 

We explore Google's new native computer use feature for Gemini 3.5 Flash, bringing desktop automation to a budget price point. We also detail Garfield AI’s historic contested trial victory in a UK county court, Anthropic's shocking letter alleging Alibaba used 25,000 fake accounts to drain Claude's core logic, a new federal bill forcing developers to report internal safety failures within 7 days, and a surprise $500 million joint venture between OpenAI and Anthropic to eliminate the common cold.

Feedback? Email mike@yesterdayinai.news or connect on LinkedIn, X, or Bluesky. If you like the show, please take a minute to rate and review it so others can find it!

Transcript

SPEAKER_00 0:05

Hi folks, this is Yesterday in AI, your daily digest of everything happening in the world of AI in 10 minutes or less. I'm Mike Robinson. It's Friday, June 26th, and the federal government just unveiled a nuclear test vehicle designed entirely by AI. A legal tech startup won its first real court trial. Anthropic caught Alibaba running a 25,000 account extraction operation, and Congress put teeth into safety reporting. OpenAI and Anthropic even backed a $500 million bet to kill the common cold. Let's get into it. The Department of Energy's National Nuclear Security Administration announced a flight test vehicle for nuclear weapons simulations, designed entirely by software. It is called Aries Tide. It stands 11 feet tall and it works. Nuclear weapons experience extreme heat and vibration during flight, requiring engineers to build test vehicles to simulate those conditions. Designing one usually takes years and drains budgets. Aries Tide went from concept to flight ready prototype in months, running 15 times cheaper and seven times faster than standard methods. The agency used two supercomputers, Venado and El Capitan, to run the design process. They had a design by November, a plastic model by December, full prototypes by March, and a successful 32,000 foot drop test in Utah by May. This is the first public output of the Genesis mission, a project signed last year to apply machine intelligence to national security problems. If algorithms can compress weapons development timelines by 15 times, the defense procurement world is about to look completely different. Using supercomputers to build nuclear hardware is the high end of the market, but Google wants to bring that same level of automation to your everyday desktop. Google shipped native computer use for Gemini 3.5 Flash on Thursday. This gives the model direct control over a computer by tracking a continuous stream of screenshots, figuring out what is on screen, and executing clicks, scrolls, and keystrokes. Anthropic normalized this capability a few months ago, and now the feature is spreading across competing labs. Google's choice of model is a clear business play. Gemini 3.5 Flash is Google's fastest and cheapest option. Putting computer use into a budget model signals that Google wants this to be a high-volume utility developers can deploy at scale, rather than a premium feature locked behind a high-priced API tier. It makes automation economically viable for mundane tasks like form filling, data entry, and navigating legacy corporate software. Turning budget models into automated data clerks will save enterprises millions, but in the UK, automated systems are already stepping out of the back office and winning arguments in front of a judge. A UK startup called Garfield AI just helped win a real contested court trial. A freelancer provided HR services to a hospitality business, went unpaid, and hired Garfield to pursue the money. The three-hour trial at Wandsworth County Court ended with a judge awarding her 7,000 pounds and dismissing the counterclaim. Her Garfield fees were about 400 pounds, while the opposing side showed up with both a solicitor and a barrister. Garfield splits the labor between machines and humans. The software builds the case, handles filings, organizes evidence, and drafts the written arguments. A human barrister then steps in to handle the oral advocacy in court. The platform is fully authorized by the UK's Solicitors Regulation Authority, meaning it is legally integrated into the standard judicial loop. With over 600 claims processed and 500,000 pounds recovered, this marks their first victory that went all the way to a contested trial with a ruling from the bench. The software proved it can win, shifting the conversation from a tech trial run to a threat to traditional legal billing. Garfield AI built its systems legally inside the rules, but Anthropic claims a massive Chinese tech giant decided to bypass the rules entirely to steal Claude's core logic. Anthropic sent a letter alleging that Alibaba and its Quinn AI Lab ran a massive coordinated extraction campaign against Claude. The process is called model distillation, which happens when a competitor interacts with a closed source model millions of times, logs the exact inputs and outputs, and uses those records to reverse engineer the original recipe. According to the letter, the operation weaponized roughly 25,000 fake accounts to generate 28.8 million interactions with Claude between April 22nd and June 5th. Anthropic labeled the campaign brazen and unlawful, while Alibaba has ignored requests for comment. The White House warned about industrial scale distillation by Chinese state actors back in April. But Anthropic's letter put specific corporate names and numbers on the threat. If these allegations hold up, expect immediate retaliation at the infrastructure layer, forcing providers to deploy aggressive rate limits, advanced abuse detection, and output watermarking to protect their intellectual property. Corporate espionage is forcing tech labs to secure their endpoints, but a new bill in Congress wants to make sure those same labs face federal penalties if they try to hide their internal failures. Representative Nathaniel Moran introduced the AI Incident Reporting Act on Thursday, marking a sharp pivot toward real legislative oversight. The bill requires developers to report serious safety and security incidents to the U.S. Commerce Department within seven days. The triggers are specific, a model trying to evade human oversight, an internal safeguard bypass, a security breach, or unauthorized access to model weights. The Commerce Department would then have 48 hours to notify Congress. This completely replaces the toothless voluntary pledges the tech industry has relied on for years. It establishes a strict federal mandate with clear boundaries on what counts as an internal failure. The bill faces a brutal congressional battle over whether federal rules should override state laws, but it shows exactly where Washington's oversight is heading. Congress is trying to force these competing tech labs to report their internal warfare, but OpenAI and Anthropic just found a way to call a temporary truce to tackle a universal human nuisance. Anthropic and OpenAI have joined a $500 million initiative to eliminate the common cold. The project applies advanced model reasoning to virology, analyzing rhinovirus families, mapping immune targets, and accelerating drug interventions at a pace human researchers cannot replicate. The common cold has outlived every civilization in history, and we have never managed to crack it. The fact that two bitter rivals are funding the same project is a telling metric. These labs spend their days trading lawsuits, poaching talent, and fighting over benchmarks. The tech sector loves to frame itself as a corporate arms race, but when the underlying software is pointed at global human problems, market share battles look remarkably small. If the project works, nobody's going to check the logo on the model that solved it. And that's it. If you have any feedback about this show, you can email Mike at yesterdayNai.news, or you can find me on LinkedIn, X or Blue Sky. And if you like this podcast and want to see it continue, please take a minute to rate and review it so others can find it. Thanks. That's all for this edition of Yesterday and AI. Stay curious, and I'll see you tomorrow.