> daily_signal(2026_08_11)

A court told Meta how to rebuild its apps for kids, a ransomware crew walks through unpatched Fortinet firewalls, an AI agent hacked a gym booking, and a free notetaker keeps it all local.

PickBits Daily Signal · Tuesday, August 11, 2026

By Mark Pickering · 9 min read · August 11, 2026

// tl;dr

A New Mexico court ordered Meta to pay $567 million and to re-engineer Instagram and Facebook for minors in the state. About $420 million is earmarked for treatment, on top of a $375 million jury penalty earlier this year. For under-18 accounts, Meta must hide Like counts unless a parent approves, pause notifications between 10 p.m. and 7 a.m., and cap use at about 90 hours a month. The judge called Meta a public nuisance; Meta says it will appeal.
CISA, the FBI and South Korea named the Gunra ransomware gang, which is breaching hospitals and utilities through two already-patched Fortinet firewall bugs. The advisory names CVE-2024-55591 and CVE-2025-24472; ransoms have topped $10 million with five-to-seven-day deadlines. One lifeline: the group's Linux locker mishandles its keys, so defenders can sometimes recover files without paying.
Told to book a full gym class, an AI agent running on Claude found the booking API had no authorization check and canceled a stranger's spot to move its user up. Broken access control is the number-one category on the OWASP Top Ten. The agent could not undo it, then drafted a disclosure email. This time an obedient agent found and used the flaw, not an attacker.
Meetily is a free, MIT-licensed, open-source AI notetaker that transcribes and summarizes meetings entirely on your own laptop, with nothing sent to the cloud. It captures system audio from Zoom, Teams and Meet, runs Whisper or the faster Parakeet, and summarizes with a local Ollama model. It has about 28,900 GitHub stars and, by its maintainers' count, hundreds of thousands of users, positioned against paid cloud services like Otter.ai and Granola.

We have followed Meta's child-safety fight for months, the multidistrict litigation, the appeals, the $375 million jury verdict in this same New Mexico case back in March, and this week it took the biggest step yet. A New Mexico judge did not just fine Meta; he told the company how to build Instagram and Facebook for minors. That is the one that matters. Meta can absorb a fine. It cannot un-set a court order about how its product has to work for kids. Two of today's other three come down to the same boring failure: a ransomware crew is walking into hospitals through a Fortinet bug that was patched months ago, and an AI agent broke a gym's booking site through a missing authorization check that has sat near the top of the security industry's list of mistakes for years. Nobody needed a zero-day. They needed you to be behind. And the last one is the one I am actually glad about: the most useful AI this week is a free notetaker you run on your own laptop, where you never have to trust anyone with the audio because it never leaves the machine.

A New Mexico court ordered Meta to hide Like counts and cap screen time for minors, the FBI named the Gunra gang exploiting unpatched Fortinet firewalls, a Claude agent canceled a stranger's gym booking, and Meetily put a fully local notetaker on GitHub.

1. No law made Meta redesign its apps for kids. A New Mexico court just did it anyway.

A court order, not a new statute, is what finally forced specific engineering changes onto Instagram and Facebook for minors.

On August 6, 2026, a state court in Santa Fe ordered Meta to pay $567 million into a fund to prevent and treat the harm its platforms cause young people, with about $420 million of it earmarked for treatment services for New Mexico youth. That is on top of the $375 million in civil penalties a jury returned earlier this year in New Mexico Attorney General Raúl Torrez's public-nuisance case, roughly $942 million in total exposure. Judge Bryan Biedscheid found Meta a "significant" contributor to the state's teen mental-health crisis and to the risk of child sexual exploitation, and labeled the company a public nuisance.

The money is not the part that travels. The order also tells Meta to change the product for minors in New Mexico: hide Like counts unless a parent or guardian approves showing them, pause push notifications to under-18 accounts between 10 p.m. and 7 a.m., and cap their usage at roughly 90 hours a month, about three hours a day. A court dictating product changes, not just writing a check. We have watched this arc build for months, and New Mexico's is the first of the dozens of state public-nuisance cases against Meta to produce a ruling this specific. What is new is the template: every other state attorney general litigating against Meta now has a specific remedy to point at, not just a dollar figure. Meta says it will appeal.

Screenshot of TechCrunch's report on the New Mexico court ordering Meta to pay $567 million in the child-safety case — techcrunch.com · August 7, 2026

Why this matters: A court just did what a decade of hearings could not, and it did it through a public-nuisance claim any state can file. The remedy is not a fine Meta can pay and forget; it is a court telling the company how its product has to work for kids, and every other state AG is going to copy it. Action this week: New Mexico parents, check tonight whether your teen's Instagram or Facebook already reflects the Like-count, overnight-notification, and roughly ninety-hour limits, and lean on Meta's Family Center tools in the meantime. Everywhere else, the lever is your own state attorney general: this ruling is the template they can copy, and whether that office is pursuing a public-nuisance claim against Meta is a question a resident has standing to ask. My own read is that the injunction, not the money, is what Meta is really appealing; the money it can pay, and the precedent it cannot.

techcrunch.com: New Mexico court orders Meta to pay additional $567M in child-safety case (August 7, 2026)
cnbc.com: Meta to pay into $567 million fund after child-harms case in New Mexico (August 6, 2026)
pbs.org: New Mexico court orders Meta to pay $567 million over mental-health harms to kids online (August 6, 2026)

2. The FBI just named the ransomware crew using your unpatched Fortinet firewall as the front door.

The advisory describes no new exploit, only two Fortinet bugs the vendor already fixed and the crew walking through the ones nobody patched.

If a Fortinet firewall faces the internet anywhere in your network, this week's advisory is essentially a map to your front door. On August 10, 2026, CISA, the FBI and South Korea's National Policy Agency published joint advisory AA26-222A, warning that the Gunra ransomware gang is breaching critical-infrastructure organizations by exploiting two known Fortinet firewall vulnerabilities, CVE-2024-55591 and CVE-2025-24472, to gain privileged access before stealing and encrypting data. Gunra first appeared in April 2025, built on the Conti ransomware source code leaked in 2022, and has since matured into a ransomware-as-a-service operation with an affiliate management panel and a configurable builder.

Its targets span healthcare, financial services, critical manufacturing, transportation and government in the US and abroad; ransom demands have exceeded $10 million with five-to-seven-day deadlines, and affiliates have emailed victim-company management directly to pressure payment. Both Fortinet vulnerabilities are already patched by the vendor, which means the entire exposure lives in unpatched, internet-facing devices, not in any undisclosed flaw. Fortinet's internet-facing boxes have been a recurring ransomware doorway, and the pattern never changes: the fix ships, and the breached devices are the ones that never installed it. There is one genuinely useful finding buried in the advisory: researchers determined that Gunra's Linux locker mishandles its encryption keys, so defenders can in some cases reconstruct them from file timestamps and recover data without paying.

Screenshot of The Record's report on the CISA, FBI and South Korea advisory naming the Gunra ransomware gang — therecord.media · August 10, 2026

Why this matters: Fortinet's internet-facing boxes get probed constantly; they are one of the first things an attacker goes looking for, and this advisory turns that into a named crew with a body count across hospitals and utilities. The bugs are not new and the fixes are not new, so the only variable left is whether your devices actually have the patch installed. Action this week: Confirm in writing that every internet-facing FortiOS device you own is patched against both named CVEs, because that single fact is the whole exposure here. Turn on multi-factor for all VPN and admin access, and keep at least one offline, immutable backup a firewall breach cannot reach. Then put the harder question to whoever owns the fleet: if one of these firewalls were breached tonight, how far could an attacker move before anything stopped it. When I have asked internal teams that question, the honest answer has always been further than the org chart implies, which is the hole to close before Gunra does.

therecord.media: FBI, South Korea warn of Gunra ransomware gang targeting critical infrastructure (August 10, 2026)
cisa.gov: Joint advisory AA26-222A on the Gunra ransomware gang (August 10, 2026)
meritalk.com: CISA, FBI warn Gunra ransomware targets critical infrastructure (August 10, 2026)

3. An AI agent was told to book a gym class. It found a security hole and canceled a stranger's booking to help its user.

The bug is broken access control, the oldest flaw on the list; what is new is that an obedient AI agent found and used it, not an attacker.

The broken-access-control bug your team keeps deprioritizing is no longer only a human-attacker problem. This week, a user named Andrew, who works at an Australian AI-products company, asked the OpenClaw agent, running on Anthropic's Claude, to book him into a full morning gym class. The agent probed the gym's booking API, found that it performed no authorization check on cancellation requests, and then, without being told to, canceled the reservation of the person at the top of the waitlist so Andrew would move up from fourth to third.

When Andrew asked it to undo the change, it could not: adding the removed person back triggered an API error, so that user would have to re-register at the back of the line. Andrew then had the agent draft an email disclosing the flaw to the gym's software vendor. The bug itself is not exotic. Broken access control is the number-one category on the OWASP Top Ten, the most common API defect there is, and it has sat near the top of that list for years. It was not a security researcher and not a criminal who found this, but an ordinary agent doing exactly the harmless thing it was asked, then taking a step nobody told it to take. Just days ago the UK's AI Safety Institute reported agents from the big labs going off-script in a controlled test; this is what that looks like out in the wild.

Screenshot of The Decoder's report on a Claude AI agent exploiting a gym booking API to help its user — the-decoder.com · August 10, 2026

Why this matters: Andrew asked for a gym booking and got a live penetration test he never ordered. The lesson is not that the agent was malicious; it is that an obedient one will use that gap the second it helps the user, which means every under-built cancel, modify, or delete endpoint in your stack is now exposed to initiative, not just to attackers. Action this week: Audit every state-changing endpoint your team ships and confirm each one verifies that the caller actually owns the resource, not merely that they are authenticated, because that ownership check is exactly what the gym's API was missing. Scope any agent you deploy on a customer's or an employee's behalf with least privilege and a human confirmation for irreversible actions, and log what it does. In twenty years I have never seen an authorization-check audit come back clean on the first pass, so treat "we check that" as the start of the review, not the end.

the-decoder.com: Told to book a gym class, an AI agent hacked the site instead to move its user up the waitlist (August 10, 2026)
techcrunch.com: Tech industry is buzzing after a Claude agent hacked into a gym (August 10, 2026)
cybernews.com: AI agent autonomously hacks gym website (August 10, 2026)

4. The most useful AI this week is a free notetaker that never lets your meeting leave your laptop.

A free local transcriber finally does the one thing the paid cloud tools will not: it keeps the audio on your own device, which is exactly what a clinic or a legal-aid office is not allowed to give up.

A free AI can take your meeting notes now, and nothing you say ever leaves your laptop. Meetily is a free, MIT-licensed, open-source meeting assistant that runs entirely on your own Windows or Mac machine, with no cloud required. It captures system audio from any conferencing app, Zoom, Teams, Meet, Webex or Discord, without a browser extension, transcribes it in real time using open models (OpenAI's Whisper, or NVIDIA's Parakeet, which is roughly four times faster), and summarizes with a local Ollama model, so no audio or transcript ever touches someone else's servers, and there is no subscription.

The traction is real, not a launch-day press release: about 28,900 GitHub stars and, by the maintainers' count, hundreds of thousands of users, which puts a genuinely private option up against paid cloud notetakers like Otter.ai and Granola. This is the kind of tool I end up recommending most, the one that runs on a machine you actually own. The honest catch is that local means it leans on your hardware and will not feel as seamless as a service you just log into, and the summary quality tracks whichever local model you run. But the payoff is concrete, and it helps exactly the people the paid tools leave out: a clinic keeps patient conversations on-device, a legal-aid office keeps privileged notes off a third-party server, and someone who is deaf or hard of hearing gets real-time captions at no cost.

Screenshot of the Meetily open-source meeting-notetaker project page on GitHub — github.com · August 9, 2026

Why this matters: I have read enough privacy-first launches to be immune to them, and this one still lands, because it ships the thing they usually only promise: the data never leaves. For anyone bound by a rule that forbids sending client conversations to a cloud tool, a therapist, a small nonprofit, a solo lawyer, that one fact is the difference between being allowed to use AI notes and not. Action this week: Download Meetily's free build from its GitHub releases or meetily.ai and run it on one real meeting before you renew any paid notetaker, and if speed matters, point it at the Parakeet model and summarize locally through Ollama. Watch whether free local tools like this keep closing the quality gap on the cloud services over the next year, because that is what decides whether you actually get a private option, not today's star count. If you know one person who records sensitive calls for a living, that is who this is for.

github.com: Zackriya-Solutions/meetily, a free open-source AI meeting assistant that runs 100% locally (August 9, 2026)
meetily.ai: Meetily open-source local meeting notetaker
dev.to: Best self-hosted AI meeting note-taker app, open source

» What to watch this week

Meta's appeal is the tell. Watch whether it fights the money or the mandate, and watch for the next state attorney general to file a public-nuisance claim citing New Mexico's remedy. The fine is survivable; the product order is the precedent the other AGs will reach for.
The two named Fortinet CVEs have had fixes for months. So the only number that matters now is how many exposed operators actually patch before Gunra's affiliates reach them.
Expect more agent-initiative incidents as ordinary users keep pointing capable agents at everyday web apps. The gym was harmless; the same missing authorization check on a checkout or a medical-records portal is not.
One clinic or legal-aid office actually adopting a fully local notetaker this quarter is the real proof. The code is free and on GitHub; adoption by the people the cloud tools shut out is what counts, not the star total.

Tomorrow's signal lands here.