> daily_signal(2026_11_14)
Nearly 70% of US agencies now run on generative AI, and the real limits are coming from union contracts, not vendors. Plus an FBI AI-attack alert and a suit to cap frontier models.
PickBits Daily Signal · Saturday, November 14, 2026
// tl;dr
- Nearly 70% of US government agencies now use generative AI, and the limits on it are coming from union contracts, not vendors. An Ernst & Young survey put the number on it; IFPTE Local 21 in Oakland won retraining instead of AI layoffs after an October strike vote, Pennsylvania guaranteed SEIU Local 668 human oversight during an OpenAI ChatGPT Enterprise pilot, and California banned "bossware" and automated firings in September.
- The FBI issued a flash alert on Artex, an open-source AI attack framework that automates intrusions "at machine speed," with some going from first contact to admin access in under three hours. It was posted to GitHub in August by an account called "Autumn-27"; Censys found more than 1,600 exposed Artex consoles online, and Microsoft tied the tool to a supply-chain attack on ConnectWise.
- Four nonprofits sued five of the biggest AI companies in San Francisco Superior Court, asking a judge to freeze development of any model more capable than today's. Public Citizen, the Heat Initiative, the Alliance for Secure AI, and the Irreplaceable Education Fund invoked California's Unfair Competition Law and a public-nuisance theory, not a new AI statute. They are not asking to shut anything down or win money, just to draw a line above what exists today.
- Researchers in China built an AI that detects citrus greening before a tree shows symptoms, hitting nearly 98% average accuracy across five severity levels. It fuses an ordinary color photo, a hyperspectral image, and a handheld chlorophyll reading, cheap sensors that beat any one of them alone. The disease is incurable and wrecked Florida's groves; the work is still lab-stage.
The workplace story moved faster than I expected. We have tracked AI worker surveillance since the spring, and I assumed the limits would keep coming from regulators. Instead they are coming from union contracts: Oakland traded layoffs for retraining, and Pennsylvania put a human sign-off into its ChatGPT pilot in writing.
The FBI put a hard clock on the hacking arc we have followed since June: a break-in that used to take a human team days can take an open-source tool a single afternoon. In San Francisco, four nonprofits are asking a judge to stop the labs shipping anything more capable than today's models. And a lab in Guangxi caught incurable citrus greening from a phone photo before a tree looks sick.
The one that stuck with me: a union local in Oakland got a human put back in the loop, in writing.
1. Unions, not vendors, are writing binding AI rules inside US government agencies.
Nearly 70% of agencies already use generative AI. The news this week is who got to write the rules on it, and it was the unions, through contracts and strike votes.
An Ernst & Young survey found that nearly 70% of US government agencies already use generative AI. The unions are now the ones setting the terms. IFPTE Local 21 in Oakland forced the city into retraining commitments instead of AI-driven layoffs after an October strike vote. Pennsylvania Governor Josh Shapiro guaranteed SEIU Local 668 human oversight during a pilot of OpenAI's ChatGPT Enterprise. A Transport Workers Union local in Ohio won veto power over autonomous vehicles.
State law is moving the same direction. In September, Newsom banned "bossware" surveillance and fully automated firings in California. New York went earlier, protecting its state employees from AI displacement by statute back in late 2025. The Oakland fight is the one that got my attention: a strike vote that turned into a retraining commitment instead of layoffs. Ohio's transit local went a completely different route and took a flat veto over self-driving vehicles.
Why this matters: A human sign-off on the AI was the kind of thing that lived in an internal policy nobody enforced. A union contract or a state law can now make you actually prove it's there. Honestly, that is good news even if you're the one deploying this: writing that step in now is cheaper than bolting it on after a grievance or a lawsuit forces it.
Action this week: Write the human-oversight step into your AI rollout now, before a union or a state law writes it for you. Say who signs off before the AI can fire or discipline someone, and how they get to push back. And if you operate in California, treat Newsom's September bossware and automated-firing ban as live legal risk, so I would check any monitoring or scheduling tool for those features before the lawyers do.
the-decoder.com: Nearly 70 percent of US agencies use generative AI, and unions are finally catching up with guardrails (November 13, 2026)
ey.com: EY state, local, and education generative-AI adoption survey (October 2026)
theguardian.com: US workers and governments turn to unions for AI guardrails (November 12, 2026)
2. The FBI says an open-source AI tool is breaking into agencies in under three hours.
An attacker can go from a first probe to owning your whole network in a single afternoon.
On November 13, the FBI issued a flash alert on Artex, an open-source AI attack framework now used by both criminals and nation-state hackers against government agencies and critical-infrastructure operators. Artex can automate intrusions "at machine speed, from reconnaissance through data exfiltration," the FBI says, with some progressing "from initial targeting to privileged access in under three hours." That used to take a human team days. The tool was posted to GitHub in August by an account called "Autumn-27"; the developer, identified by the Wall Street Journal as Chinese security engineer Li Puhua, disowned the misuse on October 8 and pulled the repository, but the FBI says it persists through forks and a commercial build.
Censys found more than 1,600 exposed Artex console instances sitting on the public internet right now, and Microsoft tied Artex to a supply-chain attack on ConnectWise. We ran the Five Eyes "months, not years" warning back in June. The FBI just put a number on it: three hours. The hole is still the one you left open: an unpatched box, an internet-facing console, an admin login without a second factor. AI just makes finding it nearly free.
Why this matters: The FBI alert is a public record you can act on, and the number in it wrecks how most teams plan a response. Three hours from first probe to domain admin means the plan that waits for someone to walk in Monday morning is already too late. Plan for a three-hour break-in. The old assumption was that you had days, because a person was doing the work by hand. Not anymore.
Action this week: Confirm none of your own management consoles face the internet, because an exposed console with a weak login is exactly the door Artex is built to walk through, and it is the cheapest check you can run. Then rehearse containment on a three-hour clock instead of a three-day one. When I have asked teams to time themselves on this, the gap is almost always who gets paged at 2 a.m., not the tools.
3. Four nonprofits are asking a California judge to freeze any AI more advanced than today's.
Rather than an AI statute, the plaintiffs reached for an ordinary consumer-protection law.
On November 12, four public-interest nonprofits (Public Citizen, the Heat Initiative, the Alliance for Secure AI, and the Irreplaceable Education Fund) filed suit in San Francisco Superior Court against OpenAI, Anthropic, Google, Meta, and two of SpaceX's AI units. It is a strange ask. They are not trying to shut anything down or win money, only to stop any model from getting more capable than each company's best one as of that date. Today's public models would stay on the market; the suit just draws a line above them.
The plaintiffs aren't citing an AI statute at all. They invoke California's Unfair Competition Law and a public-nuisance claim: building anything more capable than today's models is itself a public-safety risk, with bioweapons and large-scale cyberattacks as the danger they name. As of November 13, Reuters reported no injunction had issued and none of the five defendants had responded. We have watched suits against the AI labs build up over the past month or two. This one asks for a freeze at today's level, not a shutdown; the models already out stay out.
Why this matters: They built the case on California's consumer-protection law, not an AI statute, and asked a judge to stop the labs shipping anything more capable than today's models. Even if the freeze never issues, the next advocacy group or attorney general can file the same thing for the price of a court fee. Honestly, I don't think they win. But if the judge doesn't toss it in the first week, that alone is the story.
Action this week: Track this docket if you build on frontier models, and watch whether the judge takes the theory seriously or throws it out early. A copycat suit is cheap to file, so work out now what a capability cap at today's level would actually cost you.
unite.ai: Public Citizen, advocacy groups sue to freeze frontier AI development (November 2026)
cryptobriefing.com: Nonprofits sue AI labs to halt frontier development (November 2026)
Listen to the Daily Signal podcast at pickbits.ai/daily-signal.
4. An AI spots citrus greening before the tree looks sick, at nearly 98% accuracy.
Once a tree looks sick it is usually too late, so the only window to act is before symptoms show.
Citrus greening has no cure, and it has wrecked groves in Florida and Brazil. By the time a tree looks infected, it's usually too late to stop the spread. Researchers at China's Guilin University of Technology and the Guangxi Academy of Specialty Crops have built a system that detects the disease before visible symptoms appear. They trained it on 2,700 citrus leaves, from healthy trees up through five levels of HLB severity, including infections you cannot see yet. It hit nearly 98% average accuracy at sorting how bad each infection was.
The hardware is the surprise here. It runs on cheap gear, an ordinary color photo, a hyperspectral image, and a chlorophyll reading off a handheld SPAD meter. That combination beat plain photos or hyperspectral images used on their own. The researchers are upfront that this is still experimental (field trials, more citrus varieties, and whole-canopy assessment are all listed as next steps), so the 98% is a lab-stage result on 2,700 leaves, not a tool a grower can buy yet.
Why this matters: For a grower, the only thing that helps is catching it before symptoms, while pulling one infected tree can still save the grove around it. The gear is cheap: a color photo, a hyperspectral image, and a handheld chlorophyll reading, and together they beat any single one of them. My own read: keep an eye on this one, but nobody is buying it for a while yet, and the striking part is how little hardware it needed.
Action this week: Read the 98% as what it is: lab-stage, on 2,700 leaves, not field-tested. Watch for the field trial across more varieties before anyone calls this a tool a grower can use. If you fund agri-tech, the part worth backing is the low-cost sensor fusion: three cheap readings that beat any one of them on its own.
freshfruitportal.com: AI could detect HLB before visible symptoms, study finds (November 13, 2026)
citrusindustry.net: AI model shows promise for early HLB detection (October 30, 2026)
» What to watch this week
- Does anyone in the private sector pick up the same contract language? The IFPTE, SEIU, and TWU wins are the ones a private union would copy. I'll be watching whether a works council or a private union gets the same thing into its next contract: retraining instead of layoffs, or a person who has to sign off before the AI can act.
- How many of those 1,600+ exposed consoles come offline. That is the number that tells me whether critical-infrastructure operators actually rewrote their incident plans around a three-hour intrusion, or kept drilling for the multi-day version that is no longer the threat.
- The judge's first move on the lawsuit. No injunction has issued and no defendant has responded, so the early signal is simple: does the court take the public-nuisance theory seriously, or toss it, and does anyone file a copycat.
- Can the citrus model hold up outside the lab? The 98% is on 2,700 leaves, so I want to see the field trial, across more varieties, before calling it a tool a grower can use.
Tomorrow's signal lands here.