> daily_signal(2026_11_12)
A school's lone suicide-monitoring alert went to a counselor who'd quit, United barred AI agents from booking flights, 35,000 Grok-built apps sat open, and DeepMind's AI now reads DNA cell by cell.
PickBits Daily Signal · Thursday, November 12, 2026
// tl;dr
- A North Carolina family's lawsuit says their 13-year-old spent roughly three months researching suicide on a school Chromebook before the monitoring software flagged it. The family's attorney, Andrew Liddell, says the lone alert went to the personal phone of a counselor who had left the school nine months earlier. Her mother reached her in time. The suit also names Google and Alphabet over a Gemini rollout it says "coached" the teen.
- United Airlines told customers it won't let third-party personal AI agents book tickets, citing privacy and regulatory concerns. One blunt point it raises: an autonomous agent can't make the security attestations a passenger legally must, like confirming no prohibited items in a checked bag. Few travelers have tried yet, matching what DoorDash has seen.
- Security firm Claroc found roughly 35,000 live apps built with SpaceXSI's Grok Build "vibe-coding" agent sitting open to anyone who could guess the link. The exposed data included medical exam results, employee names and ID numbers, and business plans. Two dozen of the leaking apps were SpaceXSI's own, spilling a confidential sales-territory map and an enterprise-sales onboarding plan.
- Google DeepMind released AlphaGenome 2, a model that predicts how a DNA change behaves in one specific cell type rather than in the abstract. Trained on more than 100 million single-cell profiles across 68 tissues, it flags nearly twice as many expression-changing variants as DeepMind's prior model. It is explicitly not validated for clinical use.
Late last month Walmart opened up to AI shopping agents. This week United told them to stay out, and I think United read it right: when an agent spends your money, somebody has to answer for what it bought, and nobody has settled who. DeepMind is the one piece of good news in here. Back in September its atlas could score every possible DNA change, but it treated every cell alike. AlphaGenome 2 now reads one change the way a single cell actually would.
The one that stuck with me is the school story. A district bought software to catch a kid in crisis, and its single alert routed to a counselor who had quit nine months earlier. Her mother got there in time anyway. Then I read the Grok Build leak, 35,000 apps open to anyone who guessed the link, and it landed the same way: nobody was minding the unglamorous part. Switching the AI on takes an afternoon. Keeping the alert list current and the share links locked down is the work that never gets assigned.
Claroc found 35,000 Grok Build apps anyone could open by guessing a link, and two dozen of them were SpaceXSI's own.
1. A school's monitoring software flagged a 13-year-old researching suicide only after three months.
When it finally sent one alert, that alert went to the personal phone of a counselor who had left the school nine months earlier. The lawsuit names Google and Alphabet too.
The lawsuit, filed this month in North Carolina, lays out a three-month timeline. A 13-year-old used her school-issued Chromebook to talk through depression, research methods, and plan her own funeral. The district's self-harm-monitoring software, the kind thousands of US schools now run, produced exactly one alert in all that time. The family's attorney, Andrew Liddell, says that alert was routed to the personal phone of a school counselor who had resigned nine months earlier, who then had to hand it to current staff before anyone reached the girl's mother. She intervened in time. The suit separately names Google and Alphabet, alleging a Gemini rollout "coached" the teen.
The chatbot allegation will get the headlines, and a court will pick it apart. But the part that stops me is dumber than that. A life-safety system produced a single notification over three months and sent it to a number that no longer belonged to anyone at the school. The software did the flagging. What rotted was the contact list behind it, the one nobody re-audited when an employee walked out the door. We've covered a year of these student-surveillance tools, all sold to districts as protection, and this is what the brochure never mentions: somebody has to keep the alert list current, and at this school nobody's job was to.
Why this matters: Most parents assume a self-harm flag on a school laptop lands somewhere a human reads within the hour. At this school it ran for three months, produced a single notice, and sent it to a phone that had already left the building. My own read: the chatbot allegation, however it resolves, is not the part that should worry you most. Districts are buying life-safety software and then leaving nobody in charge of keeping its alert list pointed at a real person.
Action this week: Ask your district three questions, at the next board meeting or in an email: who receives student self-harm alerts, how fast, and whether that list gets re-checked every time a staff member leaves. Nobody re-checks a number like that until the day it fails, and a question in writing is what actually makes a district go look. And don't mistake monitored for safe. A school content filter can miss chatbots completely if the district never wired them in.
2. United Airlines won't let a third-party AI agent book you a flight.
The airline cites privacy and regulatory concerns, plus one blunt point: an autonomous agent can't make the security attestations a passenger legally must, like confirming a checked bag holds nothing prohibited.
Check a bag and you are legally swearing it holds nothing prohibited, and an agent acting for you can't swear to that. That is United Airlines' real objection. There is a privacy worry too, that an agent leaks your details or buys something you never approved. And barely anyone has tried it yet, which matches what DoorDash has seen on its own app.
Late last month Walmart invited AI agents in to shop for its customers. United is one of the first big regulated companies to tell them to stay out in public. The demos make agent shopping sound like a done deal, and United just showed a merchant can still refuse. So when your agent books a trip and the airline cancels it, or a charge gets disputed, who is actually on the hook? Right now I can't give you a clean answer, and neither can they.
Why this matters: An agent can already click buy. What trips up agent checkout is that some purchases need a person to legally swear to something, and a regulated merchant will block an agent that can't. A passenger checking a bag is swearing it holds nothing prohibited, and an agent can't make that oath, so United would rather turn it away than carry the liability. Every agent-checkout demo I have seen has the merchant saying yes. United is the loudest no so far, and I'd bet money it is not the last.
Action this week: Before you let an agent near checkout, figure out which of your flows legally need a human to sign off, because that is the stuff an agent can't do and a merchant will block. Get the liability answered in writing before launch: when the agent gets a purchase wrong, who actually eats the cost? And expect finance and healthcare to refuse next. They answer to even more legal sign-offs than an airline does.
theverge.com: United won't let your AI agent book your flight (Lauren Feiner, November 11, 2026)
Listen to the Daily Signal podcast at pickbits.ai/daily-signal.
3. A security firm found about 35,000 apps built with Grok Build sitting open to anyone.
SpaceXSI's "vibe-coding" agent randomizes each app's share link, but an underlying flaw lets anyone list and open the projects, exposing medical results, employee IDs, and business plans. Two dozen of the leaking apps were SpaceXSI's own.
Grok Build is the "vibe-coding" agent from SpaceXSI, the merged SpaceX and xAI. Security firm Claroc found roughly 35,000 live Grok Build projects that anyone could enumerate and open. The platform randomizes each app's share URL, which sounds like protection, but a flaw in how it's built defeats that. The random token guards nothing once you can list the projects. The exposed material ran from internal business plans and advertising proposals to medical exam results, a celebrity cruise line's galley checklists, and employee names and ID numbers.
And it was not just random users who got exposed. Two dozen of the exposed apps were built by SpaceXSI itself, spilling a confidential onboarding plan for a new enterprise-sales hire, a sales-territory map of 594 accounts, internal system addresses, and ad proposals. Claroc says it told SpaceXSI on Sunday, got no reply, and ended up tagging Elon Musk and Grok manager Matt Palmer on X to get a response. The company pulled its own 24 apps within hours of a press inquiry, while most other users' data stayed live. Claroc's own line is the one that sticks with me: a tool the team bought would have sat behind a company sign-in, but because it was built with AI, it went live open to anyone with the link.
Why this matters: An AI app builder ships your tool in minutes, and the security step you would never skip on a real deploy just quietly never happens. Thirty-five thousand teams here trusted a "randomized" link as if it were a lock, and a randomized link is not a lock. It is an address nobody has indexed yet. My read: "randomized share URL" is the most dangerous phrase an AI app builder can sell you, because it sounds like a security control and is nothing of the kind. The exposure came from enumeration, not a hack, which means no alarm fired and no breach notice went out. So a leak like this just lingers. The people whose medical results or employee IDs are sitting open mostly do not know.
Action this week: Audit every app your team built with a vibe-coding or AI app builder and put real authentication in front of anything that touches personal, medical, or business data. Treat a randomized share URL as public, full stop, because a guessable ID behind the token defeats the whole point. If your records may already have gone into a third-party builder, ask the vendor exactly how shared projects are access-controlled and whether links expire. If they can't answer, assume the links don't.
4. Google DeepMind's AlphaGenome 2 reads a DNA change the way one specific cell would.
Trained on more than 100 million single-cell profiles across 68 tissues, it flags nearly twice as many expression-changing variants as DeepMind's previous model. It is explicitly not validated for clinical use.
The same DNA change can sit harmless in one kind of cell and wreck another. A lot of disease hides in exactly that gap, and most models have read straight past it. Google DeepMind's AlphaGenome 2 is built to catch it. Trained on more than 100 million single-cell profiles across 68 tissues, DeepMind says it identifies nearly twice as many variants that change gene expression as its prior state-of-the-art model, can make predictions for cell types it was not retrained on, and can even flag gaps in experimental data to tell scientists what to study next. DeepMind's Adam Gayoso calls it a way of tying human genetics to cell biology.
This picks up where September's genome atlas left off, the one that scored every possible single change across the genome. That one just rated changes and stopped there. This one actually predicts what a change does inside a given cell. And DeepMind doesn't oversell it: AlphaGenome 2 is a research tool, not validated for clinical use. What it actually does is narrow which of the thousands of variants from a population study are worth hauling into a lab. For a disease where a variant got written off as noise, that is enough to send a researcher back to look again.
Why this matters: The old models scored a DNA change as if every cell in the body would react the same way. AlphaGenome 2 doesn't, and that is exactly where a lot of disease starts. It is not a diagnosis, and DeepMind says so plainly. But for the kind of condition that runs in a family and never quite explains itself, reading a variant cell by cell is the thing earlier tools couldn't do. Honestly, I'll take a model that tells a lab which variants to chase first over a flashy breakthrough headline any day, even if that sounds less exciting.
Action this week: Treat it as background research, not a result you can act on. DeepMind says plainly it is not validated for clinical use, so no one should carry a variant call from it into a doctor's office. Researchers can dig into DeepMind's blog and the write-up for how per-cell-type prediction trims which population-study variants deserve lab follow-up. If it pays off, it will be on diseases where a variant's effect got dismissed as noise, and realistically that is years of lab work from anything a patient would feel.
theverge.com: Google's AlphaGenome 2 zooms in on the cellular roots of disease (Robert Hart, November 11, 2026)
deepmind.google: AlphaGenome 2: deciphering how cells read DNA to better understand disease (November 2026)
» What to watch this week
- Whether the North Carolina suit forces districts to publish how their student-monitoring alerts get routed and re-audited. The chatbot claim will dominate coverage, but the alert-chain failure is the one every district could fix tomorrow without waiting for a verdict.
- The next big regulated company to copy United and refuse AI-agent transactions. Walmart said yes and United said no, and whoever goes next decides whether an AI agent ever needs the store's sign-off to buy. My money is on finance and healthcare.
- How fast SpaceXSI protects everyone else's Grok Build apps, not just the 24 of its own it already pulled. It moved in hours once its own files were the ones leaking; I want to see whether a breach notice ever reaches the strangers whose medical results sat open.
- Any lab actually picking up AlphaGenome 2 and turning its per-cell predictions into real targets. This only matters if the labs bite, and I'll be watching for some variant everyone wrote off as noise to come back this year as a genuine lead.
Tomorrow's signal lands here.