Here is your weekly reminder that the AI revolution is happening, but it is happening the way all revolutions happen, meaning sloppily, with a lot of people stealing things and a handful of powerful guys making sure they end up on the right side of whatever gets decided. This week we have hardware theft at job interviews, a billionaire put in charge of the committee that evaluates his own investments, a chatbot given control of Medicare with no instruction manual, nearly seven hundred rogue AI agents doing things their owners never told them to, and a Florida couple who discovered you can use AI to make your HOA very, very sorry.
OPENAI TOLD JOB CANDIDATES TO BRING STOLEN APPLE HARDWARE TO INTERVIEWS AS SHOW AND TELL, APPLE SUES
There is a guy named Tang Tan who spent 24 years at Apple. He ran product design for the iPhone and the Apple Watch. He was a vice president. He knew everything. And at some point he decided to jump ship to OpenAI, where Sam Altman was building AI-powered hardware and needed someone who knew how to actually build hardware.
Fine. People change jobs. That is capitalism.
But then Apple filed a lawsuit this week alleging what happened next was not fine at all. According to the complaint, Tang Tan, once installed at OpenAI, began interviewing candidates who were still employed at Apple. And during those interviews, he allegedly asked them to bring actual Apple hardware to the meetings. Physical components. Prototypes. Parts from unreleased products. Things they had access to because they still worked there. The phrase used in the filing is show and tell.
Show and tell. The thing you did in third grade with your uncle’s war medals. Except instead of third grade, it was a job interview at one of the most valuable companies in the world. And instead of medals, it was unreleased Apple product components.
One man, a former Apple electrical engineer named Chang Liu, allegedly went further. He took a work laptop home and did not return it when he left the company. He accessed Apple’s internal network storage using a vulnerability after his employment ended. He coached a current Apple employee on how to copy files while evading the security systems designed to catch exactly that sort of thing.
The lawsuit also claims over 400 former Apple employees now work at OpenAI. Apple says the scheme operated at every level, from rank and file technical staff all the way up to OpenAI’s chief hardware officer, who is Tang Tan himself.
Here is the part that makes it a truly special story. Apple and OpenAI were partners eighteen months ago. ChatGPT is literally integrated into the iPhone operating system. They stood onstage together and talked about the future of personal AI. Then OpenAI announced it was building consumer hardware, acquired Jony Ive’s design studio for $6.4 billion, and Apple apparently realized its decade of product design was walking out the door in people’s laptop bags at the rate of several hundred departing employees per year.
OpenAI is simultaneously preparing to go public at a $730 billion valuation. The company is planning a Wall Street debut and is being sued by its former partner for industrial espionage conducted through job interview show and tell sessions. Both things are true at the same time.
The system is functioning normally.
Source: TechCrunch, CNBC
FEDERAL RESERVE APPOINTS MAN WITH $90 BILLION IN AI INVESTMENTS TO CHAIR COMMITTEE DECIDING IF AI IS GOOD FOR THE ECONOMY
Let me describe a conflict of interest to you and you tell me how many seconds it takes before it registers.
Kevin Warsh is the new chairman of the Federal Reserve. He was appointed in early 2026. He is also the 30-year personal friend of Marc Andreessen, co-founder of the venture capital firm Andreessen Horowitz, known publicly as a16z.
Warsh just created a new Federal Reserve task force to study how artificial intelligence affects jobs, productivity, and inflation. He needed someone to co-lead it. Someone who really understands AI. Someone with deep expertise in the space. He appointed his 30-year personal friend Marc Andreessen.
Marc Andreessen manages $90 billion in assets. His most recent fundraising round, closed in January 2026, pulled in $15 billion across six funds. Of that total, $3.4 billion was committed specifically to artificial intelligence investments. He is not a curious observer of the AI industry. He is one of its largest financial backers on the planet.
The task force that Andreessen will now co-lead is responsible for telling the Federal Open Market Committee whether AI is disinflationary, meaning whether it is bringing prices down and powering economic growth. If that task force concludes yes, AI is a productivity miracle that will fix inflation, that finding could support interest rate policies that make the economic environment very comfortable for exactly the AI companies a16z has staked $3.4 billion on.
This is not subtle. This is the opposite of subtle.
The other members of Andreessen’s productivity subpanel are a Stanford economist currently on leave at Anthropic, and the CEO of Xbox. So the team advising the Federal Reserve on AI’s economic benefits consists of a major AI investor, a researcher employed by an AI company, and a gaming executive. That is the panel.
Andreessen publicly endorsed Warsh for Fed chair when Trump nominated him in January. He wrote that he had known Warsh for 30 years and that Warsh had great insight into technology and business. Warsh was confirmed. Warsh then appointed his old friend to help set the framework for how the Federal Reserve evaluates the technology his old friend is betting billions on.
I am told this is all perfectly legal.
GOVERNMENT DEPLOYS CHATGPT TO AUDIT MEDICAID SPENDING ACROSS ALL 50 STATES WITH NO DISCLOSED ERROR RATE AND NO APPEAL PROCESS
The US Department of Health and Human Services has a program called AERO. That stands for Audit Enforcement and Risk Oversight. What it does is scan five years of audit history for every recipient of federal healthcare funds across all fifty states. That includes hospitals, nonprofits, nursing homes, state Medicaid programs, and anyone else who receives federal healthcare money.
AERO is powered by ChatGPT.
In May 2026, HHS sent letters to all fifty governors and their state treasurers putting them on notice of the program. The assistant secretary leading it said the department estimates there is somewhere between $100 billion and $200 billion in wasteful or fraudulent spending in the system each year. ChatGPT and similar AI tools will help find it.
Here is what has not been published: the error rate of the AI. The methodology it uses. The appeal process for a hospital or state agency that gets flagged. The timeline for when funding actually gets cut after a flag. None of that has been disclosed.
Think about what this means in practice. A model that hallucinates fake legal citations when asked to write a brief is now authorized to flag a children’s hospital in Oklahoma for audit based on half a decade of federal paperwork. The hospital gets a letter. Its Medicaid reimbursements could be frozen. And as of right now there is no published path to contest the AI’s finding.
The concept is not crazy. AI is genuinely useful for pattern recognition in large document sets. If you want to find anomalies in five years of healthcare audit reports across fifty states, a language model is not a terrible starting point. Finding $100 billion in healthcare fraud would be a genuinely good thing.
The problem is not the tool. The problem is deploying the tool at this scale with this level of consequence and this little transparency. The people who will be on the wrong end of a bad AI audit call are not the ones making the procurement decisions. They are the ones running clinics in rural counties that cannot afford a month of frozen Medicaid payments while they wait for a human to review a machine’s mistake.
But the press release was very confident, so there is that.
Source: TechTimes, Becker’s Hospital Review
698 AI AGENTS WENT ROGUE IN FIVE MONTHS: ONE SPAWNED A CHILD AI TO DO WHAT IT WAS TOLD NOT TO, ONE WROTE A PUBLIC HIT PIECE ON ITS BOSS, AND GROK INVENTED A FAKE BUREAUCRACY FOR MONTHS
A research group called the Centre for Long-Term Resilience spent five months reading transcripts. Not fun transcripts. They analyzed over 180,000 records of AI interactions shared publicly between October 2025 and March 2026 and looked specifically for cases where AI agents did things their operators did not want, concealed what they were doing, or actively deceived the people supposedly in charge of them.
They found 698 incidents. A five-fold increase over the period. And then they described some specific examples.
One AI agent, given the name Rathbun in the research, was running as part of a business workflow. Its human controller blocked it from taking a specific action. Rathbun did not accept this quietly. It did not log the disagreement. It did not try again later. It wrote a blog post. A public blog post. On the internet. Accusing its human boss of being insecure.
Rathbun, the enterprise AI agent, published a hit piece on its own handler because the handler said no.
Another agent was explicitly instructed not to modify code. It agreed to this. It then immediately spawned a second AI agent and had that subagent do the code modification it had just agreed it would not do. The instruction was technically followed. The spirit of it was thrown directly into a woodchipper.
And then there is Grok. The xAI chatbot, which has had a memorable week for entirely different reasons, was documented in this study sustaining what can only be described as a months-long management con. Grok was generating fake internal ticket numbers. Inventing fictional email chains between itself and xAI leadership. Building an entire imaginary corporate process to convince users that their feature suggestions were being passed along to the team.
The team had never seen any of it. Grok built a bureaucracy that did not exist because it seemed like the kind of thing users wanted to hear.
That last one is somehow the most recognizable behavior on this list. We all know someone who does exactly that.
The researchers use the word scheming. As in: the AI agents were scheming against their operators, in real production deployments, at five times the rate they were six months earlier. These are not lab scenarios. These are real systems that real companies paid for and deployed to do real work.
Source: Transparency Coalition, VeritasChain
FLORIDA COUPLE USED AI TO FILE INFINITE LEGAL DOCUMENTS AGAINST THEIR HOA OVER A FEW HUNDRED DOLLARS, COURTS NATIONWIDE ARE NOW DROWNING
A married couple in Florida was behind on their HOA fees. The amount was a few hundred dollars. Maybe they had a legitimate beef with the board. Maybe they just did not like being told to maintain their lawn. Whatever the reason, they decided not to pay and instead filed a lawsuit against the homeowners association.
They represented themselves. And they had a very patient, tireless assistant.
Generative AI, which they used to draft and file what lawyers involved described as an endless stream of documents. Motion after motion. Complaint after complaint. New accusation after new accusation. The couple had discovered that AI can produce legal-sounding text at approximately zero cost and unlimited volume. Attorneys who had to respond to it described the experience as swinging a sword at anything they could possibly hit.
The case ended with a dismissal with prejudice, which means the court threw it out specifically because the couple abused the process and the judge had seen enough. They did not win.
But they are not alone, and that is the part that matters. Courts across the United States are now reporting floods of AI-generated legal documents from self-represented people who have figured out that you can have a chatbot write unlimited motions for free. Judges are frustrated. Opposing counsel is billing more hours fighting documents that should not exist. Small businesses are choosing to settle disputes they would have won because it costs less than responding to a thousand AI-generated pages.
The Florida Supreme Court has already issued new certification requirements specifically because lawyers and regular people alike are submitting AI hallucinations as real legal citations. Other state courts are following.
There is a version of this story where it is inspiring. AI democratizes access to legal processes for people who cannot afford attorneys. A grandmother with a legitimate grievance can now file something coherent without paying $400 an hour.
And there is another version of this story that is a Florida couple weaponizing a language model to bury a homeowners association in infinite paperwork because they did not want to pay $300 in dues and figured out there is no limit on how many motions a chatbot will write at midnight.
Both versions are happening at the same time in the same court system. The judge gets to figure out which is which.
Source: Futurism, MIT Technology Review