AI Next WaveWeek in Review: AI Found the Holes. A Disaster Alert Went Dark.
October 17 to October 23, 2026. A patch week shut public websites and a Japanese quake-alert service, Texas asked a court to keep kids off ChatGPT, and Medicare's AI gatekeeper got a closer look.
The Short Version
If you carried a SoftBank, Y!mobile or LINEMO phone in Japan on Friday morning, your phone stopped receiving emergency alert mail at 6:47 a.m. That's the channel that carries earthquake early warnings, tsunami warnings and evacuation notices. SoftBank says it switched the service off to fix a flaw in PHP, the software behind a large share of the world's websites, and plans to finish on Monday, October 26.
The flaw was found with help from AI. Two days earlier, PHP's developers had shipped emergency fixes, and councils, police sites, banks, railways and tax offices from Edinburgh to Tokyo went offline while they patched.
That pattern ran through the whole week. A system made a call or went quiet, and an ordinary person was left asking who answers for it: the customer waiting on an alert, the Texas parent whose child talks to a chatbot, the Medicare patient waiting on an approval.
The Patch Week
On Wednesday, PHP released fixes for two critical flaws, rated 9.2 out of 10, that could let an attacker take over a web server without logging in. Federal cybersecurity, intelligence and law enforcement agencies told organizations to update right away, or take systems offline if they couldn't. Edinburgh and Oxford council sites and the Metropolitan Police site went unavailable, Insurance Business reported, and Dutch, Finnish, French, Danish and German services went dark too.
In Japan, the list ran long. Reuters reported that Nomura, Mitsubishi UFJ, ANA, JR East, the Finance Ministry, the National Tax Agency and the Meteorological Agency all suspended some online services. The government responded by announcing a team that will use AI to simulate attacks on its own systems. "Defenders also need to make use of AI and respond proactively, staying one step ahead," said Ryosei Akazawa, the industry minister.
The AI angle sits at the start of the story. The PHP Foundation credited the security firms Calif and Wiz, along with OpenAI, and both firms have said they used AI models in their research, according to Tenable. Defense One reports OpenAI said Calif used one of its models to find a critical flaw.
Attackers moved fast. Amazon says it caught a first attempt on October 17, four days before the fixes, and that the flaws are now being broadly exploited. Still, SecurityWeek reported Friday that no successful break-in has been confirmed, and one security firm withdrew an earlier report of a success.
On Friday, PHP flagged a third, narrower issue that needs an unusual setup to bite. Its advice is a single setting change until the next regular release.
My read is that AI made the finding cheap, and it's making the attacking cheap too. On October 19, AI Next Wave covered a hacker whose agents broke into more than 100 company websites at about $50 per success. Banks are reacting on their side of the login. JPMorgan and Schwab bar outside agents that log in with a customer's password, though Schwab lets traders use certain outside AI tools for trades. Citi and Bank of America have described limits of their own, The Paypers reported, citing the Financial Times.
Kids and Chatbots Went to Court
On Friday, Texas Attorney General Ken Paxton sued OpenAI and Sam Altman in Collin County. The state says OpenAI misrepresented ChatGPT's safety and failed to warn users, including children, about its risks. It asks for an injunction that would bar minors in Texas from the product and require a court-appointed evaluator to approve new models before they reach Texans. The complaint's alleged harms include encouragement of self-harm. These are allegations and requests, and no court has ruled on them. OpenAI told the Texas Tribune it has worked with experts to build safeguards and would work with Texas on rules that apply to the whole industry, "not just one company."
A day earlier, Florida Attorney General James Uthmeier proposed a law that would bar chatbot operators from offering minors features that simulate companionship or suggest the bot is a real person, and would make app stores verify ages and get parental consent for downloads. It's a proposal for lawmakers, with no bill filed yet. "Parents deserve real control, and companies must design for safety, not addiction," he said.
The best measurement of the week came from Denmark. The children's group Børns Vilkår surveyed 2,692 sixth and ninth graders early this year. About 85 percent had used a chatbot in the past six months. Among users, 28 percent had longer personal conversations with one, a third of girls and about a fifth of boys. Nearly a third of ninth-grade users had a chatbot write messages to friends or family. Children who used chatbots for personal talks were more often lonely or low in self-esteem, though the report says that's a correlation and doesn't show the chatbot caused it. Its advice is that children be at least 15 before using commercial chatbots without an adult.
Wisconsin, which joined the group of state attorneys general watching Character.AI's bankruptcy on October 16, argued this week that the company's limits for minors could change under a new owner.
Picture a parent in Plano with a 14-year-old who uses ChatGPT for homework. No court order has changed anything yet, so the first move is a conversation. Ask what the chatbot got used for last week: homework, a worry, a text to a friend. That answer is your baseline. The Danish group suggests kids under 15 use a chatbot only with an adult beside them. ChatGPT's terms bar sharing an account, so for a child 13 or older the route OpenAI offers is a linked teen account with parental controls: quiet hours, and a notice if reviewers see a serious self-harm concern, which OpenAI says is not real-time monitoring. Try two weeks of schoolwork done side by side, and count what worked: assignments finished, and whether your child came to you with the personal stuff. A teenager may simply switch to a friend's phone, and the parent carries the extra time. You hold the authority to stop or adjust the plan, and your child's school sets its own rules. If self-harm comes up in any of those conversations, call or text 988 in the US.
Who Signs the Decision
Friday's edition followed WISeR, a Medicare pilot that uses AI-assisted review to decide in advance whether Medicare will pay for 13 services, including back injections, in Arizona, New Jersey, Ohio, Oklahoma, Texas and Washington. Records released after a lawsuit by the Electronic Frontier Foundation show two vendors denied 5,944 requests in three months, and on October 15 a senior Medicare official said, "There are a lot of things about WISeR that are not working."
That edition was corrected on three points on Saturday. Oklahoma rancher Bill Curry's ten hours on the road covered two trips in all, KFF Health News reported, so the new approval step added about five. Every decision carries a tracking number, approved or not, so a family has to ask whether the answer was yes. And CMS says reviewers typically decide within three days of receiving a request through their portal, with requests routed through Medicare's regional contractor sometimes taking longer, and an expedited review answered within two days when a delay could seriously harm the patient.
Italy went the other way on Friday for jobs. Its new law requires a person with real authority to sign every AI-assisted hiring, discipline and firing decision, and a dismissal that breaks the rule is void. The same day, China's State Council issued guidelines with 18 tasks, including a mechanism to assess job risks from AI adoption. Xinhua describes them as guidance, with no penalties named.
Power, Water and Who Pays
Microsoft confirmed on October 19 it had walked away from anchoring Nscale's planned AI data center in Loughton, England, after grid officials said power might not arrive until the 2030s, Reuters reported. Nscale says it remains committed to the site. In Mississippi, a federal court on October 19 refused the Justice Department's attempt to dismiss the NAACP's Clean Air Act suit over unpermitted gas turbines that power xAI's Colossus 2 data center, Earthjustice says. The suit alleges more than 5,000 tons of nitrogen oxides a year, and the case now moves ahead.
Status Check
PHP's fixes are out for its supported versions, and older versions need workarounds or replacement. SoftBank's alert service stays off until its planned Monday completion; its notice showed no restoration as of Saturday evening. Neither critical flaw was on the federal list of known exploited vulnerabilities as of Friday.
The Texas suit is filed and awaits OpenAI's formal answer. Florida's plan is a proposal. Character.AI's proposed sale schedule puts bids due November 20.
WISeR continues, with no full public evaluation and no published count of reversed denials.
What to Watch Next Week
On Monday, SoftBank plans to finish the maintenance and restore emergency alerts, and says it will notify customers when it does. PHP's next regular release is expected to carry the fix and full advisory for the third issue. Further out, Character.AI's proposed sale schedule has bids due November 20, and Wisconsin wants stronger user protections attached to any deal.
What You Can Do With This
If you or your parents rely on phone alerts
Check that your phone has a second route to emergency warnings, such as a local government or weather app with alerts on, and ask your county how it sends evacuation notices.
If you're a parent
Ask your child what they used a chatbot for this week, and listen for the personal uses as much as homework. For a younger child, the Danish advice is simple: use it together, with you in the room. For a teenager on ChatGPT, link your accounts through its parental controls rather than sharing your login, which the terms don't allow.
If you run a small business, school or clinic website
Ask whoever hosts your site, in writing, whether PHP was updated last week and which services went offline to do it. If a site can't be patched quickly, European government security teams advise taking it off the public internet until a fix is applied and checked, so agree on a fallback that keeps people able to reach you.
If you or a parent use traditional Medicare in the six states
Ask the doctor's office whether a planned procedure needs WISeR approval, note the date the request goes in, and ask whether the written decision says yes.
The Bigger Picture
This week showed AI cutting the cost of finding flaws for defenders and attackers at once. The fixes still ran on people's time: an IT worker taking a council site offline, a carrier switching off a quake-alert service for several days, a parent learning what a child tells a chatbot. My read is that more patch weeks like this one are coming, as AI tools comb through old software that everyone depends on. The places that come through them best will be the ones that planned a backup for the people who rely on them, before the service went quiet.
References
SoftBank: Notice on suspension of emergency alert mail for PHP vulnerability maintenance (Oct. 23, 2026) Reuters via The Star: Japan to simulate cyberattacks as PHP flaws roil global internet (Oct. 23, 2026) PHP: News archive 2026, security releases (Oct. 21) and informational advisory on CVE-2026-115560 (Oct. 23, 2026) Tenable: php2shell frequently asked questions (Oct. 21, updated Oct. 23, 2026) SecurityWeek: Mirai botnet operators join PHP2Shell attacks as ASLR bypass exploits emerge (Oct. 23, 2026) Defense One: Critical flaws in PHP prompt warning (Oct. 21, 2026) Insurance Business UK: UK councils and Met Police among sites pulled offline after NCSC alert (Oct. 21, 2026) ExecutiveGov: NSA, CISA urge immediate patching of PHP2Shell flaws (Oct. 22, 2026) CERT-EU: Security Advisory 2026-018, Critical Vulnerabilities in PHP (v1.2, Oct. 23, 2026) The Paypers: Citi and Bank of America restrict AI agents from accessing accounts (Oct. 19, 2026) American Banker: Schwab blocks unauthorized AI agents, points clients to Charley (Oct. 15, updated Oct. 16, 2026) Texas Tribune: Texas Attorney General Ken Paxton sues ChatGPT-maker OpenAI (Oct. 23, 2026) Texas Attorney General: Paxton sues OpenAI and CEO Sam Altman (Oct. 23, 2026) Courthouse News: Texas attorney general sues OpenAI over ChatGPT safety and safeguards (Oct. 23, 2026) ClickOrlando/WKMG: Florida AG proposes new law to crack down on teens, social media (Oct. 23, 2026) Børns Vilkår: Mellem mig og chatbotten, report (Oct. 2026) OpenAI: Terms of Use, Registration and access (effective Jan. 1, 2026) OpenAI Help Center: Parental controls in ChatGPT (accessed Oct. 24, 2026) UpNorthNews: Wisconsin seeks consumer protections as AI chatbot company files for bankruptcy (Oct. 23, 2026) PPC Land: Character.ai faces creditors under oath on November 19, a day before bids (Oct. 16, 2026) KFF Health News: Medicare's AI push snarls patients and doctors in errors and delays (June 23, 2026) CMS: WISeR Model Frequently Asked Questions (modified Sept. 22, 2026) CMS: WISeR Provider and Supplier Operational Guide, Version 7.0 (July 24, 2026) Electronic Frontier Foundation: New records reveal problems with Medicare's AI prior authorization experiment (Sept. 8, corrected Oct. 9, 2026) Medical Daily: CMS official admits many parts of WISeR are not working (Oct. 18, 2026) ECNews: Utilizzo di sistemi di AI in azienda, D.Lgs. 179/2026 (Oct. 13, 2026) Xinhua: China issues guidelines to promote employment amid AI advances (Oct. 23, 2026) Reuters via The Star: Microsoft pulls out of UK data center deal with Nscale (Oct. 19, 2026) Earthjustice: Judge throws out DOJ attempt to deny communities' day in court against xAI (Oct. 20, 2026)