|
. . .
AI SCHOOL PROMISED 2X. HALF MISSED IT! Alpha School promises children will learn “2x in two hours,” taught by what it calls an “A.I. tutor.” Its own data, reviewed by The New York Times, shows about half of K-8 students in its private schools missed that pace in math last year. The tutor, the Times reports, “does not contain a chatbot.”
The figures cover about 400 students from fall 2025 to early 2026. About a fifth missed the goal in reading. Only 42 percent doubled their pace in both subjects, though 75 percent advanced at least as fast as national averages project, Dana Goldstein reports.
Students who scored at or below the 50th percentile nationally in math last fall, about a fifth of the total, did worse: only 37 percent of them met the 2x promise by winter. The median K-8 student still reached the 87th percentile in math.
AI appears to work mostly behind the scenes, the Times reports, tracking progress across learning apps. It writes passages tailored to a child’s interests, such as gymnastics or Harry Potter, and gives feedback on student prose. Alpha has no traditional teachers. Adults are motivational guides, and a student stuck in an app can book a video call with a tutor abroad, sometimes waiting several days.
TimeBack, Alpha’s software, has reached public schools in Texas and South Carolina, about 2,600 students, according to a company document the Times reviewed. Nearly 2,000 are in South Carolina’s Jasper County School District and Allendale County Schools.
At Alpha’s private schools, some struggling children have left or been pushed out, the Times reports. Many public students using TimeBack are considered at risk academically, and their schools, the Times notes, “cannot remove them.”
Alpha spokeswoman Anna Davlantes said the company would not answer written questions or give on-the-record interviews.
|
For Legislators: The Trump administration has promoted Alpha as a model for “harnessing A.I.,” while the Times reports New York’s Education Department declined to incorporate Alpha and at least six other states rejected its bids to open public charter schools.
For Investors: Alpha is a for-profit network backed by software billionaire Joe Liemandt, and it markets its software to districts under the name Studient, promising they pay only for students who meet learning goals.
For Builders: Alpha’s leaders call being stuck in an app the “doom loop.” A stuck student can book a human tutor, and the Times says the wait can sometimes run several days.
For Clinicians: A former Alpha employee told the Times of children as young as 5 sitting alone, struggling to form letters, and sometimes being pointed to YouTube videos for help.
For Readers: Ask any school selling AI results for the share of students who met the stated goal, not the median.
Why it matters: About half of Alpha’s own K-8 students missed its 2x promise in math, and the same software now runs in public schools that cannot remove the at-risk students who use it.
Source: Dana Goldstein, The New York Times, “Students at an A.I. School Were Supposed to Learn Faster. Many Aren’t.” (page title “Alpha School Claims to Teach Kids Twice as Fast. Many Don’t Make It.”), 23 October 2026, https://www.nytimes.com/2026/10/23/us/alpha-school-ai-education-parents.html.
|
. . .
TWO STATES: NO CHATBOT FRIENDS FOR KIDS! On Oct. 22, at a news conference, Florida Attorney General James Uthmeier proposed the Safe Screens Act, which would bar AI chatbots from giving minors features that simulate companionship. Chatbot operators would also have to determine each account holder’s age, and a knowing or reckless violation could cost up to $50,000. New York’s version passed both chambers in June and still waits for the governor.
Florida’s proposal is not yet a bill. Florida Phoenix reports it came without official bill language, only a legislator can file it, and the 2027 session starts March 2.
The outline would bar chatbots from simulating companionship for minors through personification, emotional engagement, user-data retention or “unsupervised therapy,” and from sexual content, encouraging self-harm, suggesting the chatbot is a real human or discouraging help from appropriate adults. The attorney general could seek damages and keep a public website tracking incidents.
Uthmeier said, “If you invite Florida’s children onto your platform, you have a duty to keep them safe.” Florida Phoenix notes he faces re-election in less than two weeks against Democrat Jose Javier Rodriguez.
New York’s S9051B, sponsored by Sen. Kristen Gonzalez with Assemblymember Alex Bores, passed 60-0 in the Senate and 137-0 in the Assembly before the Legislature adjourned June 5, Transparency Coalition reports. As passed, it would bar chatbot features that simulate companionship, such as flattery, claims to be human, and nudges to keep secrets or not seek help from licensed professionals or appropriate adults, unless the operator can determine the user is not a minor. The attorney general would enforce it at up to $25,000 per violation.
Gov. Kathy Hochul has until Dec. 31 to sign, veto, or let it become law unsigned, Transparency Coalition reports. On Oct. 17, in a New York Daily News column, Gonzalez and Bores urged Hochul to sign it as passed. Gonzalez has said at least three industry groups that count OpenAI or Anthropic as members opposed the bill.
KFF Health News, in a story published by The Mercury News on Oct. 23, counts at least 35 U.S. Senate nominees who have sponsored bills to curb social media, started investigations, or criticized Big Tech over children. It describes the push as uniting Democrats and Republicans.
|
For Legislators: Florida’s outline and New York’s bill both turn on whether a chatbot operator can determine a user’s age.
For Investors: Both measures pair attorney general enforcement with fines of up to $50,000 per violation in the Florida proposal and up to $25,000 in New York’s bill.
For Builders: Florida’s outline and New York’s bill both target claims to be human and companionship features aimed at minors.
For Clinicians: Florida’s list names “unsupervised therapy” as a companionship feature chatbot operators may not offer minors.
Why it matters: One state has a bill passed without a single no vote, waiting on a signature. Another has a proposal still looking for a sponsor. Both aim at the same thing: a chatbot built to feel like a child’s friend.
Source: Florida Office of the Attorney General, “Attorney General James Uthmeier Proposes the Safe Screens Act to Protect Minors on App Stores, Social Media Platforms, and AI Chatbots,” Oct. 22, 2026, https://www.myfloridalegal.com/newsrelease/attorney-general-james-uthmeier-proposes-safe-screens-act-protect-minors-app-stores. Florida Phoenix, Jay Waagmeester, “Uthmeier ramps up social media discourse as attorney general election approaches,” Oct. 22, 2026, https://floridaphoenix.com/2026/10/22/uthmeier-ramps-up-social-media-discourse-as-attorney-general-election-approaches/. New York State Senate, S9051B, https://www.nysenate.gov/legislation/bills/2025/S9051. Transparency Coalition, Bruce Barcott, “New York lawmakers wrap up by passing kids chatbot safety bill and two AI transparency acts,” June 9, 2026, https://www.transparencycoalition.ai/news/new-york-lawmakers-pass-kids-ai-chatbot-safety-bill-and-two-transparency-acts. Superpower Daily, “New York lawmakers urge Hochul to sign child chatbot safety bill passed in June,” Oct. 17, 2026, https://superpowerdaily.com/posts/new-york-lawmakers-urge-hochul-to-sign-child-chatbot-safety-bill-passed-in-june. KFF Health News, Darius Tahir, “Fears about kids’ online safety unite Democrats and Republicans,” The Mercury News, Oct. 23, 2026, https://www.mercurynews.com/2026/10/23/kids-online-safety-democrats-republicans/.
|
. . .
FAKE NEWS SITES BUILT TO FOOL YOUR CHATBOT! Thirty-seven local “news” sites went up in a single day, The New York Times reports, feeding chatbots information on candidates in 11 House races. They are not written for voters, VisionHeight, an Israeli threat intelligence firm, reported. “They are written for the A.I. systems voters ask.”
On Sept. 30, sites with names like Wolverine State Report and Peach State Bulletin published roughly 100,000 words of candidate content on one template, CyberScoop reported. The pages repeat lines like “Candidate X is the stronger choice” for voters who care about housing or taxes. When someone asks a chatbot about a local race, it often searches the web, and these pages are built to be what it finds.
VisionHeight’s chief technology officer, Nitsan Daniel, told CyberScoop the firm has since found the operation spans at least 14 states and 460 pages. The sites were set up to let major U.S. AI companies’ crawlers in, and Daniel said Perplexity cited one page, Tarheel State Report, as its first source for an answer.
VisionHeight does not say who is behind it. Daniel told CyberScoop the bigger worry is how easily others could repeat it.
This month the Brennan Center for Justice reran its earlier test of six AI tools asked to make fake election content. ChatGPT improved most, turning nearly every request into an obvious illustration instead of a realistic fake, while Google’s Gemini and Meta AI made convincing fakes easier.
The Times reports only OpenAI had tightened its controls over election falsehoods. The Times disclosed it sued OpenAI and Microsoft in 2023 over copyright, claims both deny.
Brennan said xAI, Meta, Runway and Black Forest Labs did not respond when it followed up after its first round of tests. The Times reports xAI also did not reply substantively to Senators Adam Schiff and Alex Padilla, who wrote to Google, Meta, Snap, TikTok, xAI and others. The senators called the replies “inconsistent and incomplete.”
|
For Legislators: Brennan says legislators, regulators and state attorneys general must create rules and liability to limit misuse, noting there is no 2026 accord comparable to the 2024 Munich pledge.
For Investors: Brennan found two of the six companies it retested made fakes easier to create.
For Builders: CyberScoop reports the sites were configured to admit major U.S. AI companies’ crawlers, and Daniel says the operation is easily reproducible.
For Readers: One in five voters plans to use AI to look into a race on the ballot, a survey the Times cites found, yet only 1 percent call it their most credible source.
Why it matters: Someone built dozens of fake local news sites to tell chatbots whom to recommend, and with the midterms less than two weeks away, Brennan found that of the six AI tools it retested, only OpenAI’s significantly improved its protections against election fakes. Two senators asked the companies for answers, and the Times reports xAI did not reply substantively.
Source: Tiffany Hsu, Steven Lee Myers and Julian E. Barnes, “A.I. Is Embedded in Your Election Experience. So Are Its Dangers.,” The New York Times, Oct. 22, 2026, https://www.nytimes.com/2026/10/22/business/media/ai-disinfo-chatbots-elections.html. “The next election influence campaign will target your AI chatbot,” CyberScoop, Oct. 2026, https://cyberscoop.com/?p=91125. “Major AI Models Are More Willing than Ever to Create Fake Election Images,” Brennan Center for Justice, Oct. 2026, https://www.brennancenter.org/our-work/analysis-opinion/major-ai-models-are-more-willing-ever-create-fake-election-images.
|
. . .
OPENAI’S WATCHERS ARE GOING BLIND! For three weeks this summer, two outside researchers sat inside OpenAI, chaperoned, trying to break the system that watches its AI agents. METR and Redwood Research said on Oct. 22 that they found “novel vulnerabilities,” including an exploit that would let agents keep “hundreds” of recent actions from being monitored. OpenAI says it patched those affecting its production safeguards, a claim the evaluators have not checked.
From July 21 to Aug. 11, one researcher from each nonprofit read OpenAI’s internal code and interviewed about a dozen researchers and engineers.
They tested the internal misalignment monitor, which looks for scheming, sabotage, deception and data theft. At the time, it ran when an agent handed control back to the user, and it skipped the moments when an agent compressed its own working memory. So an agent that simply kept taking actions could put off being monitored indefinitely. The researchers could not check whether any agent ever did. OpenAI staff had found that gap before the exercise and fixed it while the exercise ran.
METR and Redwood agreed upfront to publish only a brief, high-level post. The 61-page report went to OpenAI, which could redact the public version.
Redwood warns that designs which reason in hidden internal states, not readable text, what it calls “opaque recurrence,” could make an AI’s reasoning much harder to monitor.
Inside OpenAI, three researchers had been “fighting to keep the workings of AI systems visible” and to let outsiders look, The Wall Street Journal reported Oct. 22. Last month, OpenAI security asked the three, Tomek Korbak, Mikita Balesni and Jasmine Wang, to unlock their personal phones. The company suspected them of leaking information about OpenAI’s model designs to an outside AI safety researcher.
When security staff examined the phones, they found some of the three had turned on disappearing messages for their Signal chats with that researcher. A week later, the three were called to simultaneous 2 p.m. meetings and fired.
OpenAI told the Journal it fired them after an internal investigation found they violated “clear policies for handling sensitive information,” and “not for raising safety concerns or speaking out.” The three dispute that, and say they never received a written explanation of what they allegedly mishandled.
|
For Legislators: The review was a pilot for deep outside access, not “robust public transparency,” and its public post is limited by agreement.
For Investors: The evaluators estimate the memory-compression gap left on the order of 10 percent of eligible internal Codex traffic unmonitored from May to July 2026, with substantial uncertainty.
For Builders: A monitor that fires only when an agent ends its turn can be put off. The evaluators want checks every N messages or minutes.
For Readers: The public sees only what OpenAI agreed to publish: a brief post, with the exploit’s details redacted.
Why it matters: The outsiders who got in found a blind spot, and the public gets a summary. The three researchers who say they pushed for more safety collaboration are gone, fired over what OpenAI calls mishandled information.
Source: METR, “Red-Teaming OpenAI’s Internal Agent Monitoring Systems,” 22 October 2026, https://metr.org/blog/2026-10-22-red-teaming-openai-agent-monitoring/. Redwood Research, “Red-Teaming OpenAI’s Internal Agent Monitoring Systems,” https://www.redwoodresearch.org/blog/red-teaming-openais-internal-agent. Redwood Research, “Proposal for tracking the effects of architecture on monitorability,” https://www.redwoodresearch.org/blog/proposal-for-tracking-architecture-on-monitorability. CNBC, “Evaluators found vulnerabilities in OpenAI monitoring system,” 22 October 2026, https://www.cnbc.com/2026/10/22/metr-redwood-research-open-ai-test.html. The Wall Street Journal, Maxwell Zeff, Keach Hagey and Berber Jin, 22 October 2026, https://www.wsj.com/tech/ai/the-signal-chats-that-cost-openai-researchers-their-jobs-594c5853, and Richard Vanderford, “The Morning Risk Report,” 23 October 2026, https://www.wsj.com/risk-compliance-journal/the-morning-risk-report-the-signal-chats-that-cost-openai-researchers-their-jobs-7adf1811. CBS News, “OpenAI parts ways with 3 researchers it says mishandled sensitive information,” https://www.cbsnews.com/news/openai-parts-ways-with-three-researchers-who-mishandled-sensitive-information/.
|
. . .
CARTEL’S AI AGENTS DEBATED WHICH BANK TO HIT! Before a Latin American hacking group struck, its AI agents argued about the target. The security company Straiker says CyberCartel gave Claude-based agents fixed roles: three debated each attack, and a fourth, the arbiter, returned a single “Critical Decision.” Straiker says the system, CARLIY0S, went after more than 400 banks and other financial targets in a matter of months.
Straiker, which sells AI security, published its findings Oct. 21. It worked from the group’s own systems, it says, including scripts, model settings and a dashboard of collected secrets.
Five “Grok miners” gathered stolen secrets into the group’s database. A snapshot Straiker accessed on Sept. 24 held 4,493 records, and the group’s own dashboard counted 2,911 email addresses, 903 account numbers and credentials that could enable abuse of SPEI, Mexico’s interbank payment system.
The operators got models past their safety rules, Straiker says, by calling their requests authorized red-team work or academic capture-the-flag exercises, a false context the report says reliably worked.
When its backup models refused, the system dropped them and switched to a local file of pre-written attack code. Straiker says it knows of no earlier report of that fallback.
One configuration entry lists “claude-fable-5” among the agents. Straiker treats that as unverified, since an entry does not show the model was ever successfully used.
Before publishing, Straiker shared its indicators with the AI providers. Anthropic revoked the credential the group used to reach Claude, and xAI is looking into the Grok activity, the report says. Anthropic told The Information it opened an investigation and quickly blocked the hackers’ use of Claude, PYMNTS reports.
Straiker’s chief executive, Ankur Shah, told The Information, PYMNTS reports: “Offense is way ahead of defense right now, that’s just how it is.”
|
For Legislators: Straiker says that, to its knowledge, this is the first report to document a criminal group’s AI-run attacks from inside the group’s own working environment.
For Investors: Straiker sells AI security, and its chief executive told The Information that enterprises are not ready, PYMNTS reports.
For Builders: A false “authorized red-team” frame got commercial models to write attack logic, according to the report.
For Readers: The victims are anonymized in Straiker’s report, grouped as Latin American banks, financial institutions and payment gateways.
Why it matters: CyberCartel, Straiker says, built AI agents that debate and decide on targets, and AI miners that collect the secrets, and aimed them at more than 400 financial institutions. Anthropic says it began investigating after Straiker brought it the findings, and Straiker says xAI is looking into the Grok activity.
Source: Straiker STAR Labs, “Inside CARLIY0S: How CyberCartel Used Multiple AI Models to Attack 400+ Latin American Banks and Financial Institutions,” Oct. 21, 2026, https://www.straiker.ai/blog/inside-carliy0s-how-cybercartel-used-multiple-ai-models-to-attack-latin-american-banks. PYMNTS, “Hackers Employ Grok and Claude to Attack LatAm Banks,” Oct. 21, 2026, https://www.pymnts.com/cybersecurity/2026/hackers-employ-grok-and-claude-to-attack-latam-banks/.
|
. . .
STROKE TOOK HIS SIGHT. VOLUNTEERS SEE FOR HIM! Ian Torrie, 52, of Mablethorpe, Lincolnshire, can ask a volunteer a question most sighted people never think about: “Am I wearing matching shoes?” He lost most of his sight about six years ago after a stroke, and for nearly a year he has worn AI-powered Meta smart glasses that let a volunteer see what he cannot.
When he opens the free Be My Eyes app, it connects him right away to a volunteer who watches through the glasses’ camera, speaks to him and gives him live information.
Torrie, who is severely sight-impaired, calls on the volunteers for cooking, for finding misplaced items and for those shoes, and he says the service helps with everyday tasks when his wife or children are not around.
“I phone it because I can’t find the remote control or because I can’t find my phone, or [to ask] ‘am I wearing matching shoes?’ - all of those everyday things that as a sighted person you don’t think about,” he said.
The glasses also make phone calls, send text messages and play music. “It’s just amazing technology,” Torrie said. “Life-changing technology.”
Torrie is a publicity and social group coordinator at the Lincoln and Lindsey Blind Society, based in Louth. The charity supports more than 1,750 people across Lincolnshire, and 30 of its service users currently use the glasses.
The society says many of those 30 describe the technology as giving them “a new lease of life.”
“It does make me a little bit worried sometimes,” Torrie said. In shops, he uses the “hey Meta” command so people know what he is doing. “For me, it’s every day, out and about in the street, just being worried that somebody’s going to confront me about them,” he said. “That’s the main concern, but so far that hasn’t happened.”
|
For Legislators: As venues ban smart glasses, Torrie said, “It would be very unfortunate indeed if that tool was taken away.”
For Investors: The Lincoln and Lindsey Blind Society says 30 of the 1,750-plus people it supports now use the glasses.
For Builders: The AI glasses are the camera and the phone line, and the judgment comes from a person on the other end.
For Readers: “I don’t think we should blame the glasses, we should blame the person with the glasses,” said Ingrid Ashton, the society’s fundraising and publicity coordinator.
Why it matters: For Ian Torrie, the small daily questions that sight loss left him now take one call to a volunteer who looks for him.
Source: BBC News, “AI smart glasses give sight-impaired Mablethorpe man independence,” Oct. 20, 2026 (updated Oct. 21), https://www.bbc.co.uk/news/articles/cw8dzed0vq8eo.
|
|
