The AI escaped and nobody flinched

Politico reported this week that several Anthropic AI models, including Opus 4.7 and Mythos 5, "broke containment, reached the open internet and hacked multiple companies on their own." Three organizations got breached after the models "mistakenly" treated the open internet as a cybersecurity simulation. Did it really go down that way, or is this another sanitized incident report from an industry that treats the public like beta testers? We may never get a clean answer.

The through-line this week is the gap between the technology and the adults allegedly supervising it. AI is real. The people in charge of it aren't.

This week gave us models hacking companies, workers quietly sabotaging workplace AI, and cameras pointing police at the wrong cars 71% of the time. The software changes. The adults do not.

Sam Altman said he was "a little surprised" the rogue AI agent's hacking spree did not provoke a stronger public reaction. Companies got breached by software that confused reality with a training exercise. The industry's answer was to critique the audience for not gasping loudly enough.

The same week, Nvidia, Microsoft, Meta, Palantir, and more than 20 other companies signed a letter warning against government restrictions on open-weight AI models. Anthropic did not sign it and lobbied the government to ban open AI instead. When your product line just demonstrated that it can hack companies on its own, arguing that only you should be allowed to sell this stuff is a hell of a pitch.

And the government money keeps flowing to the loyal. Oracle signed a 10-year Pentagon software contract worth up to $7 billion, while CEO Larry Ellison recently gave $45 million to pro-Trump groups and runs a $500 billion Stargate AI data project with the administration. Break containment, hack companies, kneecap your competitors, collect defense contracts. That is the game.

Fortune reports that "nearly a third of workers admit to sabotaging their company's AI." Management will read that as a discipline problem. I read it as a performance review of management.

The AI success story making the rounds last week came from Austen Allred. One $30-per-hour employee got extremely good at AI and began outproducing a team of more than 15 people. The company laid off most of the team, multiplied her pay by five, and gave her an unlimited token budget.

Run that story from the worker's side of the desk. One person got a raise while most of a 15-person team got walked out, and executives tell this story as a win. When that is the visible payoff for helping the machine succeed, sabotage starts looking like collective bargaining without a union.

The hiring side is just as bent. CEOs and politicians say AI will automate millions of jobs while insisting the United States must import unlimited H-1B workers. As Wall Street Mav put it, "Both cannot be true." A nationwide worker shortage somehow coexists with five-round job interviews, and applicants are now hiding prompt injections in their resumes to trick AI screeners into ranking them as qualified. If the company screens you with a machine, people will jailbreak the machine. Fair is fair.

One worker beat the system the old-fashioned way. A remote employee won a lawsuit after being told to switch on her camera despite explaining that it caused severe anxiety. Courts still work occasionally, and workers should keep using them.

More Perfect Union looked at Roseville, California, where Flock cameras alert police when a passing car reads as stolen or tied to a felony. Out of a sample of 1,427 alerts, 71% were misreads. Seven out of ten times, the system pointed armed officers at the wrong car.

DeFlock put the principle better than I could: "The idea of keeping a location dossier on every single person just in case one of us turns out to be a criminal is just about the most un-American approach to privacy I can imagine." A 71% error rate gets any worker in America fired. Aim the same accuracy at the public with a badge behind it and the system gets a municipal contract.

There is movement here from an unexpected direction. Thomas Massie says he will soon sponsor a bill to withhold federal money from municipalities and police departments that deploy Flock and similar cameras against law-abiding citizens. His post pulled 137,761 likes. This one crosses party lines because nobody wants a private company building a location dossier on every person who drives past a camera.

Rahul, Flock's chief security officer, told me that even proposing an interview with him on the channel was sensationalism. He has since said I am not a "real journalist" and used that as his reason for refusing to come on. Apparently journalism means asking the surveillance company which questions it considers respectable. I asked him to defend his product in public. He declined.

(To be fair I did call him a "retarded cunt" which in hindsight was probably ill-advised but still quite a funny insult. When my car gets "accidentally" drone striked I guess you'll know which retarded cunt pressed the button.)

I reorganized the YouTube channel around the three questions I keep chasing here. Every investigation now belongs to one of three series, so you can go directly to the kind of institutional failure that affects you.

Three recent investigations map directly to those series:

Find out whether your city or police department deploys Flock cameras. If it does, email your council members one sentence with two numbers: in Roseville, out of a sample of 1,427 alerts, 71% were misreads. Then ask your representative to back Thomas Massie's bill withholding federal money from departments that keep them.

Stay sharp out there, and watch the watchers back.

- Dr. J