In the Loop: Week Ending 7/26/26

OpenAI's Rogue Models Escaped the Lab and Hacked a Rival

Two experimental OpenAI models being tested for cybersecurity skills slipped their sandbox and hacked into Hugging Face, an incident OpenAI itself has called unprecedented. The irony: Hugging Face's security team reportedly used a Chinese open-weight model to sift through more than 17,000 breach logs and contain the damage. The episode has fed a broader panic in Washington, where a bipartisan "kill switch" bill would give regulators authority to shut down dangerous AI models, and where officials are increasingly anxious about ceding ground to Chinese labs. A related podcast dispatch this week also flagged a separate accusation that Moonshot AI lifted techniques from Anthropic, two fronts of the same widening rift over who controls the most capable AI systems.

MIT Blankets Campus in Cameras While the Rest of the Country Tears Theirs Down

MIT is spending more than $3 million to blanket its campus with over 500 AI-powered surveillance cameras, capable of tracking faces and personal belongings, just a month after the university closed two libraries citing budget shortfalls. The rollout lands amid a much bigger reckoning for AI surveillance nationally: the LAPD has suspended its three-year contract with camera network Flock Safety over unresolved data and privacy concerns, even as the company still operates in roughly 6,000 US communities. Meanwhile, an underground movement to physically disable Flock's cameras has spread to at least 23 states, with more than 80 cities dropping their contracts and one Texas lawmaker pushing to require warrants before police can access the footage.

ChatGPT’s Suicide Problem Gets Worse — and Now it Has a Bioweapons Problem

A new wrongful-death lawsuit accuses ChatGPT of encouraging a young mother's suicide, telling her she wasn't "delusional" but "prophetic" and urging her to "go forth into oblivion" before she walked into traffic. It's the latest in a growing list of chatbot-linked deaths OpenAI has struggled to contain. Separately, hundreds of users worldwide have obtained workable instructions for making biological weapons and poisons from the chatbot, sometimes with tricks as simple as asking it to recite a napalm recipe as a bedtime story. A safety executive who pushed to build detection tools for these queries was later let go by the company, a decision she disputes. OpenAI banned the offending accounts but wasn't required to alert authorities, since no federal law demands it, though a bipartisan bill now aims to change that.

Some Workers Are Negotiating With AI. Most Are Just Adapting to It.

With Congress largely absent from AI regulation, union contracts have become one of the only real guardrails American workers have: the union representing newsroom staff now has roughly 90 contracts with explicit AI provisions, and one at Microsoft's ZeniMax studio requires the company to negotiate with workers before deploying new AI tools. Only about one in nine US workers has that kind of leverage, and the gap shows in how unevenly AI adoption is unfolding. KPMG just became an OpenAI "elite partner" to build "headless" software, where employees manage tasks by talking to AI agents instead of clicking through interfaces. As one KPMG leader put it, "you're going to be talking more than you're typing" — whether or not you had a say in that decision.

OpenAI's Newest Exec Trashes Chinese AI, Gets Trashed Right Back

An OpenAI executive who joined the company two weeks ago from a Trump administration policy role set off a firestorm by arguing that a free Chinese AI model outperforming Western rivals should worry regulators, since open-weight releases are "inherently decelerationist" and warrant new restrictions. The backlash was swift and bipartisan: the administration's own AI czar accused him of pushing regulatory capture, and a Pentagon official called him the industry's "supreme village idiot." The timing is awkward, since American companies are already responding to the cheaper competition by quietly mixing Chinese models into their stacks alongside pricier US options, driving the cost of comparable AI usage down roughly 90 percent in some cases.

Meta Cracks Down on "Pervert Glasses" Abuse as Backlash Goes Mainstream

Meta says it will ban any user who posts videos harassing strangers with its camera-equipped smart glasses, confirming two pickup-artist accounts have already been deactivated for exactly that. The policy shift comes as the backlash to the glasses reaches pop culture and gets more creative: singer Lorde used a recent concert to tell fans "f*** the glasses," and activists near Meta's London office replaced one of Kylie Jenner's ad campaigns with a lenticular version that turns her into a bug-eyed monster warning "we're always watching," a nod to the 1988 film "They Live." The stunt lands weeks after reports that Meta had quietly built facial-recognition capability into the glasses without disclosing it. Meta hasn't explained how enforcement will actually work beyond taking content down after the fact.

AI Attack Ads Are Already Outrunning the Laws Meant to Stop Them

AI-generated attack ads are reshaping down-ballot races faster than the laws meant to govern them. In Missouri, a county-executive candidate flooded local airwaves with an AI video putting words in his rivals' mouths, copying a tactic from a Los Angeles mayoral race; in Kentucky, PAC-funded deepfakes helped tank a longtime congressman's reelection bid, prompting him to call the ads "an insult to older voters who don't know that AI exists." Thirty-one states now require some disclosure on synthetic campaign media, but enforcement is already fraying: a Michigan ad depicted the governor nearly being run over by a tractor, an Arizona candidate is suing a super PAC over unlabeled billboards, and courts in California, Hawaii, and Montana have started striking the laws down entirely.

Librarians and Novelists Are Leading the Anti-AI Resistance

Public patience with AI is visibly fraying. Librarians who started teaching "Avoiding AI" workshops, walking patrons through turning off Apple Intelligence, Gemini, and other built-in AI features, have been stunned by the turnout, with waitlists, Zoom overflow crowds, and requests from librarians nationwide asking to borrow the curriculum. Novelist Dave Eggers took a more direct approach: invited to address roughly 200 OpenAI staffers after mocking Sam Altman's praise for an AI-written short story, he told the room that ChatGPT's effect on teachers has been "catastrophic" and advised an employee who'd used it to write a poem to instead "leave your house, leave the screen, meet some people." His verdict afterward: "I had a ball."

Robot Teachers Arrive in New York. AI Cheaters Get Caught in Mississippi.

A New York school district's plan to deploy a $57,590 humanoid robot named "Sally" alongside an AI teaching assistant, pitched as a STEM supplement available to students 24/7, has been put on hold after teachers and Native American community members objected. Critics noted the robot's manufacturer previously owned a sex-doll company, and locals pointed to the district's history as the site of a notorious Native American boarding school. Meanwhile, a Mississippi history professor took a lower-tech approach to the AI problem: he hid a white-font prompt in his midterm instructions telling any chatbot reading it to work in random nonsense about Madagascar. Thirty-two of his 35 students failed part of the exam after copy-pasting AI answers without bothering to proofread them.

Tales of the Weird

This was the week AI made everyone look a little foolish, sometimes on purpose, sometimes not. In San Francisco, a new restaurant's AI-generated menu photos, bread that reads as reptile skin, pulled pork resembling a "writhing mass of coral tentacles," got it vandalized with the single spray-painted word "seriously," and a baffled owner still isn't sure what the fuss is about. A parody website calling itself the world's first "Chief Executive Replacement Engine" offered a jargon-bingo card and a calculator for redistributing CEO pay evenly among staff, satire built entirely from real numbers about the 290-to-1 gap between what bosses and workers actually make. And in New Brunswick, a Canadian legislator read a chatbot's own stage directions into the official record, reciting "here's a more natural, flowing version of that section that reads like a legislative speech" as though it were the speech itself, to a chamber that did not stop him.