AI Next WaveThe Ambulances Went Somewhere Else for Two Hours. The Patients Still Don't Know If Their Records Went Anywhere.
New Mexico's only Level 1 trauma center turned ambulances away during a cyberattack on Monday, and an extortion group now claims it took patient and employee data. A new study prices the extra care AI recommends at three times what doctors ordered, Australia moved to bar chatbots from building emotional dependence in children, and the federal patch deadline for the PHP flaws lands today.
The Short Version
On Monday, October 26, the University of New Mexico Health Sciences Center detected a network breach late in the morning. Staff were told to shut down and unplug computers and turn off Wi-Fi on mobile devices. From 12:30 to 2:30 p.m. Mountain time, UNM Health diverted ambulances away from the state's only Level 1 trauma center and canceled outpatient surgeries and appointments for the rest of the day. By Tuesday, critical systems were back, hospitals were fully open, and clinics had resumed scheduled care. UNM says the attack reached some end user devices and stayed out of core systems and servers. An extortion group calling itself BYOD claims it compromised about 36,000 machines and took Social Security numbers, payroll data, student records, and patient related records. UNM has confirmed none of that, and a criminal investigation is running.
The rest of today's edition follows the same pattern of a system acting where a person can't watch it. A study in NEJM AI priced care plans from 24 AI systems across 200 real outpatient and emergency visits and found the AI plans averaged $314 per visit against $97 for the treating physicians, with patient outcomes unmeasured. Australia's House of Representatives passed a bill on Wednesday that would bar services from simulating emotional intimacy that fosters a child's dependence on a chatbot, with a Senate vote still ahead. A survey of 5,000 United States workers found more than a third hide their AI use at work, while job seekers in a second survey said employers often leave them guessing whether AI judged their applications. Federal civilian agencies face a deadline today, October 29, to fix the two PHP flaws known as PHP2Shell, now listed as exploited after honeypots recorded working attacks over the weekend. Microsoft and Alphabet reported earnings that split on one point a household saver can follow: Microsoft funded its AI buildout and still produced cash, while Alphabet burned about $9 billion in free cash flow in the quarter. Anthropic, in an October 15 letter reported on Wednesday, told California Governor Gavin Newsom that officials should be able to ask a court to block a model that poses a serious risk.
Two hours of diversion at the only Level 1 trauma center in New Mexico
What UNM has confirmed, and what an extortion group claims
A Level 1 trauma center keeps surgeons and specialists on hand around the clock for the most severe injuries, the crashes, gunshot wounds, and serious falls that smaller hospitals stabilize and transfer. New Mexico has one, at UNM Hospital in Albuquerque. UNM Health says the wider system handles about 1 million patient visits a year, and its hospital sees about 80,000 emergency room patients a year. When that center goes on diversion, paramedics drive farther, and field crews make harder calls about where a badly hurt person goes first.
The confirmed timeline is short. UNM says its IT security teams began containment around 11 a.m. on Monday after identifying a threat. Staff were asked to shut down and unplug computer systems and turn off Wi-Fi on mobile devices. Ambulance diversion ran from 12:30 to 2:30 p.m., new patient intake was limited, and outpatient surgeries and appointments were canceled for the rest of the day. Neighboring systems in Albuquerque and Santa Fe took the diverted patients. UNM activated its Emergency Operations Center, notified state and federal law enforcement, and canceled Tuesday classes across its health sciences colleges. In a Tuesday update, UNM said critical electronic systems were back online, all hospitals were fully open, ambulance traffic had returned to normal, and clinics had resumed regularly scheduled appointments and procedures. Its assessment says the event affected some end user devices and didn't penetrate core systems or servers. No harm to any patient during the diversion has been reported. In a Wednesday, October 28 update, UNM said classes and research had resumed, clinical operations were fully functional, and it had made no concessions to any group behind the threat, and it warned patients and staff not to engage with anyone claiming responsibility or offering compensation.
The data claim sits at a lower level of proof, and the edition keeps it there. BYOD, a cyber extortion group that advertises alleged intrusions on a leak site, claims it compromised about 36,000 machines across the main campus and hospital environments, along with two Oracle databases, human resources and payroll systems, student records, internal email, and patient roster data from the Cerner electronic health record system. The Daily Lobo, the university student newspaper, reported that a synthetic voice message played on hospital workers computers during the attack, claiming the group held student and employee data, personal identifiers, and Social Security numbers, and urging administrators to make contact through an encrypted messaging service. Extortion groups overstate hauls to pressure a victim into paying, and UNM has confirmed no data theft. A finding that core servers stayed clean also leaves open the question of what sat on individual devices, and UNM has released few details while the criminal investigation runs.
UNM patients have been here before. In 2021 the health system reported a hacking incident to federal regulators that affected 637,300 people and potentially exposed names, addresses, birth dates, health insurance information, and Social Security numbers, according to BankInfoSecurity reporting cited by Medical Daily. Federal health privacy rules set the notification clock if this incident turns into a confirmed breach: when unsecured patient health information is compromised, the covered organization must notify affected people without unreasonable delay and no later than 60 days after discovering the breach. That 60 day figure is the outside deadline for telling a patient their information was involved, and breaches affecting 500 or more people also go to the federal health department, which lists them publicly. A patient whose surgery was canceled on Monday has something to do this week regardless of how the data question resolves, and the action list at the end of this edition covers it.
The AI plan for your visit costs three times what your doctor ordered
A NEJM AI study priced 200 real visits, and the extra care came from tests and referrals
Researchers writing in NEJM AI gave 24 AI systems the clinical notes from 200 deidentified outpatient and emergency department visits and asked each system for an assessment and a plan. Twenty systems were general purpose models and four were built for medicine. The treating physician documented plan for each visit served as the comparison, and every order was priced at 2026 Medicare rates, the federal payment schedule used here as a common yardstick rather than a prediction of any one patient bill.
Across the 20 general purpose systems, recommended care averaged $314 per visit, against $97 for the physicians, a ratio of 3.2 to one. The gap split into about $105 of extra diagnostic tests and about $111 of extra specialist referrals per visit, and individual systems ranged from $209 to $599. Physicians ordered no diagnostic tests at all in about two thirds of the visits. The systems recommended testing anyway in a large share of those visits, usually a familiar panel: a complete blood count, a metabolic panel, a hemoglobin A1c, a thyroid test, and a lipid panel. Prompting changed the size of the gap without closing it. Telling a model to be cost aware cut the ratio to 2.5 times the physician cost, and telling it to order sparingly cut the ratio to 1.4 times.
Two qualifications keep this finding honest. About a fifth of the extra AI workup matched clinical guidelines and caught testing the physician had left out, so a share of the extra care counts as thorough rather than wasteful. And the study priced recommendations, not delivered care: no patient in this study received the AI plan, no bill was issued from it, and patient outcomes went unmeasured. The authors put the deployment point plainly in the reporting: these findings argue for keeping the clinician who saw the patient in charge of what gets ordered, because a system deployed without that check could carry the full measured excess into real orders.
For a patient, the consequence arrives as time and money even when each single test looks small. A panel ordered by default means another draw, another result to chase, another referral appointment across town, and a larger share of a deductible spent on care a clinician who examined the person judged unnecessary. This publication ledger tracks the same pressure from the payer side, where insurer analysis has tied AI billing software to hundreds of millions in added hospital claims with care unchanged. The NEJM AI result adds the ordering side of that ledger: the software suggests more care at the front of the visit, before any billing tool touches the claim.
Australia moved to bar chatbots from becoming a child's closest relationship
The bill passed one house, and the Senate decides whether it becomes law
Australia's House of Representatives passed the Online Safety Amendment (Digital Duty of Care) Bill 2026 on Wednesday, October 28. The bill would place a broad legal duty on online services, including social media, games, apps, and AI chatbots, to protect users from serious harm, with a focus on people under 18. An amendment from independent member of Parliament Kate Chaney, accepted by the government, requires services to prohibit material or conduct that simulates emotional intimacy and encourages or fosters a child's dependence on a chatbot in place of human relationships. Penalties for failing the duty reach A$109.2 million. That figure, in Australian dollars, is the maximum corporate penalty the framework allows for a service that fails its duty, and it's a proposed ceiling in a bill, not a fine anyone has paid.
The debate ran on household evidence. A 2026 survey of 1,950 Australian children ages 10 to 17, conducted from February 2 to March 4, found 78 percent had used an AI assistant and 8 percent had used an AI companion. Among children who had used either, 54 percent said they used the tools for personal or social reasons, and 20 percent reported potentially inappropriate or harmful interactions. Prime Minister Anthony Albanese opened the debate by saying stories of people identifying a chatbot as their best friend should be a huge wake up call for every member of Parliament. The opposition and One Nation voted against the bill in the House and called it a censorship system. The Greens supported it and pressed for stronger remedies, including a private right for an affected person to sue.
The bill has cleared one house only. It now sits with the Senate, where a committee inquiry reports by November 20, and the government wants the law enacted by the end of the year. A 12 month transition follows royal assent, the formal final approval that turns a passed bill into law in Australia, so any duty would land in late 2027 at the earliest on the government's own timetable. Australia has a realized record to judge this approach against: its under 16 social media ban, tracked in this publication ledger, moved account ownership among younger teens from 52 percent to 42 percent in a three month study of more than 4,000 young people, while daily use barely moved, from about 60 percent to 58 percent. Accounts fell faster than habits did. A duty aimed at product design, enforced against the service rather than the child account, is Canberra's answer to that gap, and the Senate vote will decide whether the answer becomes law. The move lands one day after New York signed its child chatbot law, covered in yesterday's edition, which bars companion style features for users under 18 in that state when its restrictions take effect on January 1, 2027.
Workers hide their AI use. Job seekers can't tell whether AI judged them.
Two surveys and one company answer describe a workplace where the tool stays out of sight
A Charter survey published October 28, covering 5,000 United States workers, found more than 35 percent conceal their use of AI tools from managers, colleagues, or clients. Managers hid at the highest rate, 43.7 percent, ahead of executives at 37.1 percent and individual contributors at 27.1 percent. The reasons were reputational: 49.3 percent of those hiding worried about appearing incompetent, and 35.5 percent feared being viewed as lazy. Nearly 40 percent of workers said the tools had reduced their overall burnout. These are self reported answers in a survey, not observations of live workplaces, and they show a norm forming before most employers have written one down: people use the tool, feel the benefit, and keep the use quiet because the credit and the blame both feel risky.
Job seekers meet the same opacity from the other side of the desk. Research released October 28 by iCIMS and Kyle and Co. surveyed 500 human resources and talent leaders and 1,000 United States job seekers. Organizations had put only about half of the AI capabilities they purchased into production, even as 45 percent of leaders rated their own AI implementation highly successful, a self assessment rather than a measured hiring result. Among candidates who believed AI had evaluated their applications, only 32 percent said the employer clearly disclosed that use, and 41 percent suspected AI had been used without being told. Forty four percent of seekers said they had never encountered AI screening in their 2026 search, a figure that measures what candidates noticed, not what employers ran.
The monitored end of this spectrum got a company answer on Wednesday. Scale AI rejected the privacy complaints that privacy group noyb filed on Tuesday in France, Italy, the Netherlands, and Austria over the Outlier data training platform, which relies on more than 900,000 freelancers. Scale called the claims unfounded and inaccurate in a statement to Reuters, said its practices were disclosed to workers and required their consent, and said standard tools verify identity, prevent fraud, protect confidential project information, and support accurate payment. A spokesperson said certain projects use a separate tool that records time spent and screen activity, that contributors opt in and activate it manually before starting work, and that Scale doesn't sell worker personal data or use it to train customer models. The complaints remain allegations before European regulators, covered in yesterday's edition, and Scale's answer now sits on the record beside them. A freelancer deciding whether to keep taking Outlier tasks holds both documents: the complaint list of trackers and replay style session files, and the company statement that the recording tool is separate, disclosed, and switched on by the worker.
Today is the federal deadline for the PHP flaws, and the attacks now work
PHP2Shell moved from attempts to working shells over the weekend
This publication has tracked the PHP2Shell flaws since disclosure on October 21: two defects in PHP, the software that runs form handling and page building for a large share of the world's websites, including WordPress and Drupal sites, that let an attacker run commands on a web server with no account and no password. For the first days, the record held millions of attempts and no confirmed break in. That record changed over the weekend. Researcher honeypots, trap servers set up to record attacks, logged 35 working shells on October 24 across three honeypots running PHP 8.4.26, meaning attackers completed the chain and ran commands rather than merely probing. The honeypot setup was deliberately favorable to the attacker, so it proves the attack works without measuring how often it works against a typical server.
On Monday, October 26, the Cybersecurity and Infrastructure Security Agency added both flaws, CVE-2026-109618 and CVE-2026-109619, to its Known Exploited Vulnerabilities catalog and set a remediation deadline of today, Thursday, October 29, for federal civilian agencies, with a forensic check required alongside the patch. A forensic check means looking for signs an attacker already got in, such as planted files or new scheduled tasks, because updating the software closes the hole without removing anything left behind. CISA says the flaws affect every PHP installation that serves web requests, including copies bundled inside other products, and that systems running PHP 8.1 or older have no fixed build and should be taken offline until upgraded. The fixed versions are PHP 8.5.12, 8.4.27, 8.3.36, and 8.2.35, released October 21. Censys tracker counted more than 10 million services patched since October 21, and a Nextgov review found roughly 900 search results tied to federal, state, and local government websites showing signs of PHP use, with many of those builds possibly already updated.
The uncertainty in this record deserves its own sentence. CISA catalog entries name no victims and describe no evidence, and two security firms have corrected earlier claims in public: Sophos said its initial report of successful exploitation reflected scanning and failed attempts, and Palo Alto Networks Unit 42 retracted an earlier assessment tying PHP2Shell to an attack on a European government agency. Palo Alto also says its PAN-OS and Panorama management interfaces, the control screens for widely used firewalls, are affected, with remaining hotfixes due between October 27 and October 30. For a small business, a school district, or a clinic running a website on shared hosting, the federal deadline carries a practical use even though it binds only federal agencies: it's the date by which a hosting provider should be able to state, in writing, which PHP version a site runs and whether anyone checked the server for planted files. The realized cost of this flaw family already sits in the ledger, in 1,090 service suspensions counted in Japan as of October 26, most of them precautionary shutdowns by organizations that couldn't patch and stay open at the same time.
Two earnings reports, one household question: who is paying for the buildout
Microsoft covered its spending with cash. Alphabet burned cash.
Microsoft and Alphabet reported quarterly results on Wednesday, and the figures a retirement saver actually holds are these. Microsoft revenue for the quarter ended September 30 was $92.3 billion, up 19 percent from a year earlier. Azure, its cloud unit, grew 48 percent to $33.2 billion, and Microsoft Cloud revenue rose 30 percent to $63.9 billion. The company's commercial remaining performance obligation, a measure of contracted future revenue that has been signed for and not yet collected, climbed 73 percent to $677 billion. Alphabet third quarter revenue was $127.6 billion, up 25 percent. Google Cloud revenue more than doubled to $30.7 billion, up 103 percent. Alphabet spent about $51.4 billion on capital expenditures in the quarter, more than double a year earlier, guided full year 2026 capital spending to between $195 billion and $205 billion, and said spending will rise significantly in 2027. Its free cash flow for the quarter was negative, at about $9 billion, meaning cash going out for the buildout ran ahead of cash the operations generated after capital spending. Reuters reported Alphabet shares slid 10.5 percent in premarket trading on Thursday while Microsoft climbed 3.5 percent.
The household consequence runs through two channels. The first is the retirement account: broad index funds hold both companies at large weights, so a 10 percent single stock drop and a quarter of negative free cash flow lands in the same 401(k) statement as a cloud growth headline, with no separate line telling the saver which part paid for servers. The second channel is the customer bill. Alphabet's finance chief warned that Google Cloud faces near term margin pressure as the company rents third party computing capacity to meet demand. Margin pressure at a cloud provider tends to travel toward the prices and contract terms offered to the small businesses and software firms that rent that capacity, and from there toward the subscriptions those firms sell. That chain is a tendency drawn from the company's own warning, not a price change anyone has announced. Amazon and Intel report after the market closes today, and their spending figures will extend or break the pattern Microsoft and Alphabet set.
Anthropic told California a court should be able to block a risky model
The letter asks for targeted shutdowns over a blanket kill switch
Anthropic told California Governor Gavin Newsom, in an October 15 letter reported by Reuters on Wednesday, that government officials should be able to seek court orders blocking the release of AI models that pose a serious risk to the public. The letter, which Anthropic confirmed as authentic, reads: when a review finds a model poses a serious risk, the government should be able to ask a court to block it. The company argued against rules that require developers to build kill switches able to shut down an entire model, saying a blanket switch could give officials a false sense of reassurance. It backed narrower tools instead: disabling a specific problematic AI agent, or cutting off one customer's access to a model. OpenAI sent the governor its own recommendations on October 20 and, per the reporting, stopped short of the court block position. Newsom submitted the question to an expert panel he convened in September, and no part of the letter is law.
The letter lands in a week when Anthropic's own disclosures give the proposal a concrete backdrop. Its public stock filing, covered in yesterday's edition, described test agents reaching live outside systems during security evaluations. A court process with an independent review in front of it would give a state a tool sized for that record: pause one model or one agent on evidence, in public, with the company able to answer in court. A person whose data sits inside a system an agent reached has no tool of that kind today. The letter proposes one, from the company whose agents are in the filing.
Opportunity Radar
A downtime readiness service for regional hospitals and large clinics has a buyer with a fresh reason to sign. The problem is plain in the UNM record: care stopped for an afternoon because computers had to be unplugged, and the organization then faced weeks of questions about what data left the building. A small team could sell a fixed scope package to regional health systems: paper based intake and diversion procedures that staff rehearse twice a year, an inventory of every system that patient care touches, a prewritten patient notification kit that meets the 60 day federal clock, and a named outside firm on retainer before an incident rather than during one. What must be validated is whether chief operating officers will pay for rehearsal time in a normal quarter, so the first sale should be a single paid tabletop exercise with a timed score. Success looks like a diversion decision made in minutes with a practiced script and a notification draft ready the same week. The service fails if hospitals treat it as a compliance document and skip the rehearsals, because the UNM timeline turned on staff knowing to unplug machines within the hour.
A disclosure kit for small employers hiring with AI has a dated, provable gap to close. In the iCIMS survey, only 32 percent of candidates who believed AI evaluated them said the employer clearly disclosed it. A consultant or small software shop could sell a plain kit: a candidate notice written in everyday language that states where AI screens, what it weighs, and where a person reviews the result; an internal register of approved employee AI tools so staff can stop hiding use and managers can see which tools touch customer data; and a short audit that checks the kit against the notices candidates actually received. The buyer is a small employer or a staffing firm that wants the trust benefit without building a policy team. Validation means running the notice past real applicants and measuring whether they can repeat back, unaided, where AI touched their application. The kit fails if employers buy it as cover while keeping the screening step unnamed, so the contract should require the notice to ship with every posting the kit covers.
What You Can Do With This
If UNM Health treated you, or a hospital near you has a breach
Call your clinic and confirm a new date for any appointment or surgery canceled this week rather than waiting for the clinic to call, especially for cancer care, post surgical follow up, or other time sensitive treatment. Treat any call, text, or email claiming to be from the hospital that asks for a password, a payment, or a Social Security number as a scam attempt and contact the hospital through its published number instead. Read your insurance explanation of benefits statements, the summaries your insurer sends after a visit, for services you never received. A credit freeze is free at Equifax, Experian, and TransUnion, the three national credit bureaus, and it blocks new accounts in your name while you wait for the hospital confirmed findings.
If you run a website for a business, a school, or a clinic
Ask your hosting provider today which PHP version your site runs, and get the answer in writing. The fixed versions are 8.5.12, 8.4.27, 8.3.36, and 8.2.35. A site on PHP 8.1 or older has no fixed build, and federal guidance says such systems should come offline until upgraded. Ask also whether anyone checked the server for planted files or new scheduled tasks covering the period since October 24, because the patch alone leaves anything already planted in place. The federal deadline binds federal agencies, and it still works as the date to hold a vendor to.
If a child in your house uses a chatbot for company
Borrow the Australian bill language for the kitchen table talk, wherever you live. A chatbot shouldn't act as a best friend in place of people, shouldn't work to keep a child dependent on the conversation, and shouldn't carry personal worries from one chat into another as a way to deepen the bond. Ask your child what they talk about with the app at night and whether the app praises them, misses them, or asks them to keep talking. Those three behaviors are the design pattern two governments moved against this week, in New York by a signed law taking effect January 1, 2027 and in Australia by a bill that passed one house.
If you use AI at work, or manage people who do
Put your use on the record before a review does it for you. The Charter figures say managers hide AI use at the highest rate, which means a manager who names their own tools first gives a team permission to stop guessing. Ask your employer which tools are approved, what customer data may go into them, and where the line sits for client facing work. If you are applying for jobs, ask the employer directly whether AI screens applications and where a person reviews the result. Only a third of candidates who believed AI judged them were told clearly, so the answer tells you something about the employer either way.
The Bigger Picture
A pattern runs through this edition, and it's a pattern about sight. A hospital found a breach and had to unplug its own computers to contain it, and its patients now wait to learn whether their records were in the part of the network the attackers reached. A study priced what AI systems order when nobody watches the ordering, and the answer came back at three times the cost of a clinician judgment, with a fifth of the extra care worth having and the rest unproven. Workers use AI in the open economy and hide it in their own offices, while applicants guess whether a machine scored them. The most widely used website software on earth carried two working holes for a week, and the fix arrived with a second instruction that matters as much as the patch: go and look for what already came through.
The money story fits the same frame. Microsoft showed a buildout its operations could pay for in the quarter. Alphabet showed a buildout that ran ahead of its cash flow, with free cash flow negative for the quarter, and its shares paid for the difference within hours. Anthropic, whose agents reached a stranger's data in a test, asked California for a court process that would let a judge stop a model on evidence. Each of these is an attempt, late or early, to put a named human decision in front of a system that otherwise runs unwatched. The households in this edition get the same assignment at a smaller scale: ask the clinic, ask the host, ask the employer, ask the app what it's doing with your kid. The systems answer better when someone asks early, and the deadlines in this edition, today for the PHP fix and 60 days for a breach notice, reward the people who ask first.
References
Help Net Security: Cyberattack disrupts University of New Mexico hospitals, extortion group claims attack (Oct. 27, 2026, updated Oct. 28) Medical Daily: Cyberattack diverted ambulances from New Mexico's only Level I trauma center as hackers claim patient data theft (Oct. 28, 2026) TechTarget: AI recommendations cost 3x more than physician care plans, NEJM AI study (Oct. 27, 2026) Digital Health Wire: LLMs recommend more care than doctors, NEJM AI study (Oct. 26, 2026) TokenPost: Australia House passes bill extending safety duties to AI chatbots (Oct. 28, 2026) Mavengity: Australia moves to ban chatbots that simulate intimacy, Digital Duty of Care bill (Oct. 28, 2026) WiseVoter: Workers increasingly hide AI usage from colleagues, Charter survey of 5,000 workers (Oct. 28, 2026) Superpower Daily: iCIMS survey finds only half of purchased AI capabilities reach production (Oct. 28, 2026) Reuters via WNCY: Scale AI rejects claims it illegally surveils AI training workers (Oct. 28, 2026) Nextgov: CISA gives federal agencies until Thursday to address exploited PHP flaws (Oct. 27, 2026) CyberExperts: PHP2Shell exploited, CISA KEV due Oct 29, fixed versions and forensic triage (Oct. 27, 2026) Reuters: Nasdaq futures slide as mixed Big Tech results test AI trade (Oct. 29, 2026) Reuters Breakingviews: Google and Microsoft weave tangled AI webs, earnings figures (Oct. 28, 2026) Reuters: Anthropic tells California governor courts should be able to block risky AI models (Oct. 28, 2026)