Issue 33 — 2026-06-14
Anthropic cuts access to Fable 5 and Mythos 5
On Friday evening, the government told Anthropic to block access to both Fable 5 and Mythos 5 models for all foreign users — inside and outside the US. The order even reached Anthropic's own employees. The company has now cut off the models for everyone.
What's missing is the 'why.' Anthropic says the government didn't spell out its national security concern. Evidence of potential jailbreaks was shared verbally, not in writing. The vulnerabilities they found were minor, and you can work around them with other models.
"They didn't say why," Anthropic told the Verge. "They just said no."
Why this matters for us: when a government order can shut down access to the tools we're building with without saying why, la gente are left guessing.
State AGs are investigating OpenAI
State attorneys general have opened an investigation into OpenAI, though it's not yet clear exactly which states are involved. The AGs are asking about OpenAI's ad policies, how it handles health data, and what it's doing more broadly with user information. The scope is broad enough to suggest regulators are looking at OpenAI as a company, not just one product.
OpenAI has been expanding beyond chatbots into advertising and health-related features. State regulators have been watching, and this looks like a coordinated effort to understand what OpenAI is doing with the data it collects and how it's using it. It's a natural next step after years of the company building products that touch more parts of people's daily lives.
Why this matters for us: if OpenAI's data practices get tightened, it could mean real changes to the tools la gente use every day — from customer service bots to health apps — and state regulators are the ones who'll hold them accountable.
Prompt injection: when AI follows the wrong instructions
An AI model is supposed to follow instructions — it's what you ask it to do. But if you feed it new instructions while it's working, it can get confused. It starts following the new ones instead.
That's prompt injection. Someone slips in a command, and the AI obeys it.
Think of it like when you're at a family gathering and tía keeps giving you new directions. She tells you to wash the dishes. Then your primo says the grill is burning. You stop washing and run to the grill. Now tía's instructions are lost.
AI does the same thing. Feed a new command into the middle of an existing conversation, and it can switch tracks.
This matters because AI apps are getting used for real work — customer service, document reading, scheduling. If the app is reading your documents and someone slips in a hidden instruction, the AI could act on that hidden instruction. It might change a setting, send an email, or approve something you didn't actually approve.
The problem isn't that AI is dumb. It's that AI follows instructions too literally. It treats every piece of text the same way — whether it came from you or from a stranger.
How do you protect yourself? Read the full message before you click, especially when it's from an unfamiliar source. Don't trust the preview. The hidden instruction is usually buried inside the text, not standing out.
When in doubt, open it in a new tab and read the whole thing before you let AI do the work.
Deal reversals don't happen every day. When they do, something's changed at the top.
— techcrunch.com
#meta-to-unwind-2b-manus-deal-after-beijing-steps-in-cbdd28A Court Says Google Is Liable for AI's False Statements
A federal court has ruled that Google must take legal responsibility for damages caused by false statements generated by its AI Overviews feature. The case involved a woman who sued after an AI-generated snippet gave her incorrect medical advice. The court's decision means Google can't dodge liability by claiming the AI simply spit out a plausible-sounding answer.
The ruling establishes that any company that designs, trains, operates, and manages an AI system is on the hook for what it produces.
TradeWork: Jobs, crews, invoices — all in one truck app
Painters, plumbers, electricians, landscapers — you know the drill. Estimates buried in your Notes app. Invoices as PDFs forwarded to your cousin's mom. Crew scheduling over text chains that stretch into oblivion. Payments coming by check, Venmo, cash — always with a reminder.
The CRMs we're sold are built for people who sit at desks. They don't fit the truck.
TradeWork is built for crews who work out of the truck. Jobs, crews, invoices, and payments in one place. Mobile-first. Bilingual surfaces so the office, the foreman, and the helper all read the same job without guessing.
Why this matters for us: when your crew can actually see what's going on without digging through text messages, you get paid faster and stop losing work to miscommunication.
https://tradework.work
Amazon's CEO Raised Concerns Before Anthropic Cut Off Two Models
Andy Jassy, Amazon's CEO, reportedly flagged security concerns about Anthropic's AI models before the government stepped in to crack down. The result: Anthropic shut off worldwide access to two of its models on Friday.
What happened here is simple but telling. Jassy, who sits on the board of Anthropic and whose company is deeply invested in the model maker, raised questions about how safe or reliable the models were. Those concerns helped set the stage for the government's intervention. Now Anthropic's models are temporarily unavailable to users around the world.
AI is building apps for regular people now
A writer at The Verge sent Gemini a prompt about her garden and came back five minutes later to find a working app in her browser window. There was also a bug. The error said, "Channel is unrecoverably broken and will be disposed!" She clicked a button. In 233 seconds Gemini reported it was done, using words like "blockages" and "race conditions." She didn't understand either one. But the app worked.
This is the new wave of "vibecoding" — building functional apps with plain English prompts instead of lines of code. The tool catches its own mistakes and asks you to approve fixes. You don't need to know how it works, just what you want it to do. The writer says this was her second or third attempt. That's the thing about AI tools: they're good enough now that you can iterate fast enough to get close, even if you're not a developer.
For working families, this is quietly important. The cousin who runs a side business, the auntie organizing her neighborhood's Facebook group, the parent juggling schedules — they've been waiting for tools that don't require a CS degree or a budget for a dev. When you can describe what you need in your own words and get something working, the barrier to entry drops. The app might not be perfect, but it's yours.
Why this matters for us: AI is becoming a tool that regular people can use to solve real problems without needing to learn how it works — and that means the next generation of apps will be built by us, not just for us.
Microsoft is weighing a spin-off for Xbox
Microsoft is preparing to lay off a significant chunk of its Xbox division and is reevaluating plans for its next-generation Project Helix console. A report from The Information says the company is considering dramatic restructuring — turning Xbox into a wholly owned subsidiary, a joint venture, or spinning it off entirely. They might even sell the business.
The report doesn't say anything is imminent. Microsoft is just laying the groundwork. But the options suggest the company is looking at Xbox differently now, trying to make the gaming division more sustainable after years of burning cash on studios and subscriptions.
The FBI built a fake town in Alabama to train for cyberattacks
The FBI has built its own replica small town inside a building in Alabama — a dedicated space for simulating real-world cyberattacks. Not a fancy lab, not a virtual environment. A physical town with streets, buildings, and infrastructure the bureau can mess with, monitor, and learn from.
It's a quieter move than the usual headline-grabbing tech news. But it speaks to something real: the FBI is building the kind of training ground it needs to understand how attacks actually unfold when they hit the places where people live and work. This isn't about training hackers. It's about training the people who respond when the lights go out.
Why this matters for us: as the FBI scales up its cyber capabilities from this fake town, the tools, protocols, and priorities that emerge will shape how federal law enforcement interacts with communities — and how quickly the bureau can act when real threats hit our streets.
KPMG yankó su reporte sobre el uso de IA por alucinaciones
KPMG sacó y volvió a meter — una y otra vez — un informe sobre cómo las empresas están usando inteligencia artificial, y los números se le vinieron abajo. TechCrunch reportó que el reporte tenía lo que se conoce como "alucinaciones": datos inventados o mal citados, el tipo de errores que nos suenan familiares porque son los mismos que cometen los modelos de lenguaje cuando se sienten confiantes y se equivocan.
No es que la IA no sirva. Es que la IA está demostrando ser una fuente poco confiable de información sobre sí misma.
Anthropic just shut the door. India's AI dreams are knocking
Anthropic paused access to its newest models, and suddenly India's tech leaders are asking whether their AI ambitions are built on sand or steel. The move ripples through a country that's betting big on artificial intelligence as a growth engine.
The question isn't just about which company you're using. It's about who controls the foundation layers that everyone else builds on. For India, where the tech ecosystem is already dense with developers, startups, and outsourced services, losing access to cutting-edge models means recalculating how much of their future depends on someone else's platform.
Tech leaders are calling it a wake-up call. The debate centers on whether India can build its own models, or if it'll keep being the world's most efficient AI customer. Either way, the clock is ticking.
Why this matters for us: When the gatekeepers of AI technology shift their rules, the communities that built their livelihoods on those platforms feel it first — and the ones who adapt fastest get to write the next chapter.
EcoFlow's PowerOcean Battery: Why Your Electricity Bill Might Finally Make Sense
Simon Hill at Wired spent time with EcoFlow's PowerOcean home battery and found it can halve his electricity bill. The battery does three jobs: it stores solar energy you already have, it buys power when it's cheap, and it keeps your lights on when the grid drops out.
The PowerOcean is a home battery — not a generator, not a solar panel, just a big box that charges and discharges. That matters because it gives you control over when you use power, not just how much you use. If you have solar panels, it catches what you make and uses it later. If you don't, it lets you buy off-peak and store it for when rates spike.
Una caja sellada de Super Mario Bros. va a $3M
Una copia de Super Mario Bros. aún en su caja y sellada con su sticker original se vendió en Heritage Auctions por $3 millones. Esto le rompe el récord anterior de $2 millones, también de una copia sellada, que se hizo en 2021.
Lo que le da su valor a esta copia en particular es que, en vez de plástico encogido (shrink wrap), estaba sellada con un sticker brillante. La casa de subastas dice que es la copia sellada más antigua que se conoce, de la segunda corrida de 1989. El sticker fue discontinuado poco después.
Y sí, esto sí salió en el bundle gratuito con la consola por $150, primo.
Why this matters for us: Los coleccionistas que pagan millones por nostalgia están apostando fuerte al valor cultural de lo que crecimos jugando — y lo que hace una pieza rara no es solo que sea vieja, sino que sea una de las primeras en su tipo.
Jane Street's formal methods, explained
Jane Street, the quantitative trading firm that runs its core systems in OCaml, has written up how formal methods work in their production code. The short version: they're using mathematical proofs and theorem provers to verify that their code does what it claims before it ships. Not unit tests. Not property-based testing. Actual proofs that the logic holds.
The post walks through how they got formal methods into a trading environment where a bug doesn't just crash a service—it costs money. They built tooling to make proof-writing feel less like writing a dissertation and more like writing code.
Para la comunidad
Tech affecting the Hispanic community
The stories below land different for our gente — immigration tech, language access, the unbanked, kids of color, gig-worker rights.
The FCC Wants to Kill Burner Phones
The FCC is proposing to strip burner phones of their anonymity. Under the plan, prepaid SIM cards — the kind you grab at 7-Eleven without a contract or credit check — would have to register with the commission and link to a real identity. The rule would require carriers to collect names, addresses, and other identifying data, then keep it for five years.
It's a quiet change with outsized consequences for la comunidad. Burner phones have long been how Brown and Black working families stay connected: no contracts, no credit checks, no long-term commitments. Undocumented parents use them. Construction workers rely on them. Abuelas get them when they want a phone that just works. The FCC's proposal would force many of those folks through a new registration process — more forms, more data, more ways for la migra and other agencies to track who's who.
The tech story here is straightforward. The FCC's original SIM registration system, created during the Obama administration, has been a mess. The new rules are meant to clean it up. But the cleanup is really a consolidation: fewer anonymous lines, more tied to real names and addresses, more exposed to surveillance. For working people who've been using prepaid phones to dodge the whole contract ecosystem, this is a direct hit.
Why this matters for us: the FCC wants to make burner phones a thing of the past, but they're erasing the tool that keeps our communities connected without forcing us into contracts, credit checks, and surveillance.
Amazon's AI Security Research Led to the White House Blocking Foreign Nationals From Anthropic's Models
Amazon's cybersecurity team spent months probing Anthropic's Fable 5 model. They fed it prompts until it started spitting out details useful for cyberattacks — things like vulnerability data that could be used against U.S. companies. They put it in a paper.
Andy Jassy took those findings to the White House. Within days, the government issued an export control directive that told Anthropic to cut off foreign nationals from using Fable 5 and Mythos 5. Anthropic complied. Amazon hasn't said anything since.
The directive didn't just hit researchers and students — it affects immigrants, workers, and small business owners who rely on these models for their work. The White House is now deciding which of us can use AI and which of us can't.
Why this matters for us: when policy makers start drawing lines around who gets access to the tools that power the economy, our comunidad ends up paying the price without always knowing who drew the line.