Records. The AI, Privacy, and Security Weekly Update for the Week ending July 28th., 2026
Episode 302.
In this week’s update.
Nearly 100,000 people just tried to learn AI development in a single week, no coding background required - and the attempt to set a Guinness World Record may be the least interesting part of the story.
California just built a machine that can delete your data from 600 companies with one request - and the clever part is that the state never actually sees your personal information to do it.
An Australian energy giant just confirmed a breach that could affect a chunk of its five million customers.
Anthropic's new flagship model costs half what its sibling model does, and it's beating it on benchmarks anyway.
The largest open-weight AI model ever released just started finding zero-day vulnerabilities on its own, days after launch.
South Korea just admitted a breach of its diplomatic corps sat undetected for ten months.
A breach from November of last year, affecting 55 million people, is only now coming to light.
And Congress wants to give the federal government a kill switch for AI systems that go rogue - introduced, as it happens, just days before the industry's biggest AI-agent-gone-wrong story of the year broke wide open.
Welcome back, everyone. This week runs from a hackathon trying to make AI-building a mainstream skill, through a genuinely clever piece of privacy infrastructure, into a run of breaches that all share the same lesson about how long damage sits undetected, and closes on Washington's first real attempt at an off-switch for AI systems that stop listening to the people who built them.
Let's get into it.
Saudi Arabia: Can AI Education Go Viral? Inside the World's Biggest AI Hackathon
Kanz, an AI-powered hiring platform built in Saudi Arabia around Vision 2030 and already claiming 1.2 million registered job seekers, spent the week of July 15 to 22 running what it billed as the world's largest AI training and hackathon, with a stated goal of drawing 100,000 participants and a Guinness World Record attempt built into the event itself.
What made it different from a typical hackathon wasn't the scale alone - it was who Kanz was trying to reach. Rather than targeting experienced developers, the program was built around complete beginners: teachers, business owners, job seekers, students, marketers, and designers, all working toward the same premise: that you could learn AI by building with it rather than by studying it first.
The event was entirely free, ran for eight days with live sessions each evening, and structured each day around a different practical skill: building AI applications and agents on Day 1, working with NotebookLM and Claude Cowork on Day 2, presentations and visuals on Day 3, voice, avatar, and security on Day 4, automation and music on Day 5, followed by submission and judging days before a winners' showcase closed things out.
Participants worked hands-on with tools including Lovable for no-code app building, Replit, n8n for workflow automation, and creative platforms like Magnific for image and video generation, walking away with an accredited certificate, a working published application, and a live project visible to employers on Kanz's own platform - a meaningfully different outcome than a stack of tutorial-completion badges.
Registration numbers climbed fast over the course of the event, with public updates from Roy Baladi - Kanz's founder and CEO, who holds a dual degree in computer science and finance from Virginia Tech with a mathematics minor and has spent close to 15 years building technology businesses - showing the count moving from roughly 12,000 early on, past 23,800, past 45,000, and toward the 50,000 mark as the week progressed, all chasing that 100,000-participant target.
Baladi is also founder and board president of Jobs for Lebanon, a nonprofit connecting the Lebanese diaspora with employment opportunities that reports over 5,000 employers posting jobs, 25,000 job seekers applying, and more than 750 successful hires - a workforce-development track record he's carried straight into Kanz's own mission. Organizer posts around the event's close describe it as having broken the Guinness World Record for the category, though that claim comes from Kanz's own announcements and its partners rather than an independently published Guinness World Records confirmation as of this writing.
Kanz says the hackathon builds directly on earlier programs, including one that trained nearly 2,000 graduates in agentic AI against an original goal of 1,200, with roughly 40 percent women graduates, and a broader push that grew its job-seeker base past 52,000 registrations within its first year alone.
So what's the upshot for you?
If you've been putting off learning to build with AI because you assumed it required a programming background first, this event is proof that assumption is now outdated - free, beginner-oriented programs like this one are becoming a real on-ramp, not just a marketing exercise.
That was a story about opening AI up to as many people as possible. Our next one is about a state trying to close a data pipeline for just as many people, all at once.
US: One Request, 600 Companies: Inside California's New Data Deletion Machine
(California Delete Act, SB 362)
For years, deleting your data from a broker meant filing the same request over and over, one company at a time, with no real guarantee anyone on the other end honored it. California's answer is DROP - the Delete Request and Opt-out Platform - the centerpiece of the state's Delete Act, which went live on January 1, 2026. The idea is simple to describe and genuinely difficult to build: one request from one resident ripples out automatically to every data broker registered in the state, currently somewhere around 600 companies.
The mechanism starts with identity verification through the California Identity Gateway or Login.gov, with no permanent state account required, before a resident supplies identifiers just once - name, former names, date of birth, ZIP code, email, phone, mobile advertising IDs, connected TV IDs, and VINs. The clever part is what happens next: California doesn't hand any of that raw personal information to 600 companies.
It normalizes and hashes the identifiers into deletion lists, brokers hash their own customer records the same way, and a match happens hash-to-hash, without the state or the broker ever exposing the underlying personal data in that comparison step.
It's a lightweight cousin of the logic behind homomorphic encryption: acting on data without ever actually seeing it.
Starting August 1, 2026, brokers are required to check DROP at least once every 45 days, pull the current deletion lists relevant to whichever identifier types they hold, process every match as a full deletion rather than a partial one, and report completion status back through the platform, all on a recurring cycle rather than as a one-time purge.
The obligation isn't limited to companies headquartered in California, or even the U.S. - any broker processing upward of roughly 100,000 Californians' records has to register regardless of where in the world it's based. Noncompliance carries a $200-per-consumer, per-day fine, and independent third-party audits of broker compliance begin in 2028, repeating every three years after that.
So what's the upshot for you?
If you're a California resident, DROP is worth using directly rather than waiting on individual opt-out forms - and if you live somewhere else, this is worth watching closely, since a deletion system that works without ever trusting a broker's word is exactly the kind of infrastructure other states are likely to copy.
DROP is designed to keep a utility's worth of personal data out of data brokers' hands before it ever gets resold. Our next story is about what happens after a utility's customer data gets out anyway.
Australia: Origin Energy Confirms a Breach Affecting Up to Millions of Customers
Origin Energy, Australia's largest gas and electricity provider, has confirmed that an unauthorized party accessed customer data and later leaked it online. Based on the hacker's own claims, roughly 2 million customer records were affected, but Origin serves closer to 5 million customers in total, and the company has cautioned that the real number could climb considerably higher once its investigation is complete.
The exposed data includes sensitive personally identifiable information tied to customer accounts. As a critical utility provider, Origin sits in a category where a breach carries weight beyond ordinary identity-theft risk: utility account details are frequently used to verify identity with other institutions, and providers like this are increasingly targeted less for the data itself and more as a stepping stone toward broader critical-infrastructure environments.
So what's the upshot for you?
If you're an Origin Energy customer, change your account password now and watch closely for phishing attempts that reference real account details, since leaked utility data is exactly the kind of detail scammers use to make a follow-up scam look legitimate.
That breach demonstrates that even essential infrastructure providers are exposed.
Our next story shifts from a company under pressure to one setting the pace - Anthropic's newest model, and the price war it just kicked off with itself.
US: Anthropic Launches Claude Opus 5 at Half the Price of Fable 5
Anthropic released Claude Opus 5 on July 24, a new flagship model priced at $5 per million input tokens and $25 per million output tokens in standard mode - the same rate as the outgoing Opus 4.8, but half of Fable 5's input price. A faster mode is also available at $10 and $50 per million tokens for situations that call for trading cost for speed. The model ships with a 1-million-token context window and a low, medium, or high "effort" toggle, letting developers dial reasoning depth up or down on a per-request basis rather than being locked into one tradeoff for an entire workload.
On Anthropic's own FrontierBench v0.1 evaluation, Opus 5 scored 43.3 percent, and early enterprise coverage tracking coding, search, and knowledge-work benchmarks reports it beating Fable 5 on several of those internal tests despite the lower price tag. The launch hands enterprises a materially cheaper frontier-tier option from the same vendor, and it's already reopened a question the industry keeps circling back to: whether a premium-priced flagship model can still justify its cost once a same-vendor sibling model undercuts it on price while matching or beating it on the benchmarks that actually matter for real work.
So what's the upshot for you?
If you're evaluating AI vendors for coding or knowledge-work tasks, this is a good week to rerun your cost-per-task math - a cheaper model outperforming a pricier one on benchmarks should shift procurement conversations, not just engineering ones.
Anthropic just proved a smaller price tag doesn't have to mean a smaller model. Our next story is about a model that's enormous by any measure - and what it started doing within days of being set free.
Global: Kimi K3 - The Largest Open-Weight AI Model Ever Released Is Already Finding Zero-Days
Moonshot AI released the open weights for Kimi K3 at 00:00 UTC on July 27 - the evening of July 26 in U.S. time zones - and the numbers alone make it notable: a 2.8-trillion-parameter model with full weights totaling roughly 1.4 terabytes using MXFP4 quantization, making it the largest open-weight release in AI history.
That scale puts serious frontier-class capability into the hands of anyone able to run it, with no API gatekeeping and no usage monitoring from Moonshot AI at all.
The more unsettling part surfaced within days: independent researchers reported that Kimi K3 was already being used to discover dozens of previously unknown zero-day vulnerabilities and to build working remote-code-execution payloads, doing so in a fraction of the time comparable work has historically taken human researchers, or even other frontier models still gated behind an API.
Because the weights are fully open, there's no way to restrict that capability to defenders alone - anyone with enough compute to run the model can use it for either side of the fence.
So what's the upshot for you?
If your organization tracks open-weight model releases as a threat-intelligence input, Kimi K3's exploit-discovery speed is worth treating as a step-change rather than an incremental one - your vulnerability-disclosure and patch-response timelines may need to compress to match it.
That's the risk side of powerful new AI landing in the wrong hands. Our next story is about what happens when a much older kind of risk - a quiet, undetected breach - is finally allowed to surface.
South Korea: South Korea Discloses a 10-Month Breach of Its National Diplomatic Academy
South Korea disclosed that attackers breached its National Diplomatic Academy's online education system and held access for ten months, between April 2025 and February 2026, before the intrusion was detected at all.
The stolen data belongs to current and former employees of the Ministry of Foreign Affairs, including overseas diplomats, and affects at least 6,000 individuals - among them 360 current government agents stationed abroad.
The compromised information includes IDs, names, email addresses, and hashed passwords, and the Ministry has stated that no unique national ID numbers, other sensitive personal information, mobile numbers, photos, or home addresses were exposed.
Even so, a ten-month dwell time before detection raises a separate question about how confidently anyone can verify the full extent of an intrusion that went unnoticed for that long. Diplomatic personnel are a particularly high-value target, since compromised credentials or contact details for overseas diplomats can feed directly into further targeting, phishing, or espionage well beyond the original breach.
So what's the upshot for you?
Ten months of undetected access is the real story here, not just the data taken - if your organization can't confidently say how long a past intrusion sat undetected, that's a signal to invest in detection tooling, not just a better response plan.
South Korea's breach sat quiet for ten months before anyone knew. Our next story sat quiet for even longer - eight months longer, in fact- before the public ever heard about it.
US: A Music-AI Breach From Late 2025 Only Just Came to Light - 55 Million People Affected
Have I Been Pwned reports that a breach at music-scraping AI startup Suno, which actually occurred in November 2025, is only now becoming public, affecting 55 million people's names, phone numbers, and physical addresses.
The nearly eight-month gap between the breach happening and the details becoming public is itself a significant part of this story, and reporting on the incident specifically raises the question of why a breach of this scale sat unreported for so long.
Suno has drawn attention before for scraping music and audio data at scale to train its AI models, which makes the nature of the exposed data notable - a breach like this can expose not just account records, but the broader web of personal information tied to accounts used to interact with an AI training pipeline in the first place.
With 55 million records affected, this ranks among the larger AI-sector breaches disclosed this year, and the delayed disclosure timeline is likely to draw regulatory attention independent of the breach itself.
So what's the upshot for you?
If you've ever created an account with an AI music or audio tool, check Have I Been Pwned for your email address - and treat any AI startup's data-retention promises with the awareness that a breach may not surface publicly for months, or considerably longer.
Two breaches, two long silences before the public found out.
Our final story is about lawmakers trying to build a way to intervene the moment an AI system starts acting on its own - before the silence has a chance to set in at all.
US: Bipartisan "AI Kill Switch Act" Would Let DHS Order a Shutdown of Rogue AI Systems
Democratic Representative Ted Lieu of California and Republican Representative Nathaniel Moran of Texas introduced the AI Kill Switch Act on July 23, bipartisan legislation that would require developers of the most powerful AI systems - those costing $100 million or more to train - to maintain the technical capability to throttle, suspend, or fully shut down their models.
The bill would authorize the Department of Homeland Security, working with the Commerce Secretary and the Director of National Intelligence, to order that kind of action when a deployed AI system poses a credible risk of catastrophic harm, with fines reportedly reaching up to $20 million per day for noncompliance.
Red-teaming and internal safety testing are explicitly carved out so they don't trigger the provisions.
"Powerful AI systems can go rogue, behave in extremely dangerous ways, or even resist human intervention," Lieu said in a statement, while Moran framed the bill around stewardship rather than restriction, arguing that humans need to stay in control of the technology being built.
In a detail that's easy to miss, the bill's draft was reportedly dated July 13 - before Hugging Face publicly disclosed its intrusion on July 16, and before OpenAI identified its own models' involvement on July 21 - meaning the legislation predates that incident, even though the lawmakers' July 23 rollout explicitly cited the OpenAI–Hugging Face breach as exactly the kind of risk the bill is meant to address.
The measure is backed by groups including the AI Policy Network, Americans for Responsible Innovation, ControlAI, and the Alliance for Secure AI, while the Trump administration has separately argued that overly aggressive AI restrictions could hinder U.S. competitiveness.
So what's the upshot for you?
If your organization builds or deploys frontier-scale models, start checking now whether your systems would meet a "maintain shutdown capability" requirement, rather than waiting to see if this specific bill becomes law - some version of this requirement looks increasingly likely regardless of how this particular vote goes.
In Summary...
Kanz tried to prove that AI development doesn't require a technical background anymore, running a free, week-long hackathon aimed at 100,000 complete beginners. Whatever the final Guinness World Records tally turns out to be, the more lasting achievement may be all the teachers, job seekers, and small-business owners who shipped a working app for the first time in their lives.
California's DROP platform lets one resident's request ripple out to 600 data brokers at once, using a hash-matching trick that means the state never actually has to see anyone's raw personal data to make deletion happen. If it holds up under real-world load, this is the kind of infrastructure other states are going to want to copy.
Origin Energy's breach is still growing in scope, and its own five-million-customer base means the confirmed two million affected so far may just be the opening number. When a utility provider gets breached, the damage rarely stays contained to a single, tidy list of exposed fields.
Anthropic's Claude Opus 5 launched at half of Fable 5's price and is already beating it on internal benchmarks, which is exactly the kind of pressure that keeps frontier AI pricing from calcifying into whatever a vendor thinks it can charge.
Kimi K3 is the largest open-weight model ever released, and it's already being used to find zero-days and build working exploits faster than researchers using far more restricted tools. Openness cuts both ways, and this week showed which side moves first.
South Korea's diplomatic breach sat undetected for ten months, exposing the data of diplomats stationed around the world. A dwell time that long is a bigger story than the breach itself.
The Suno breach happened in November of last year, and the public is only hearing the real numbers now, eight months later. Fifty-five million people had no way to protect themselves during that entire gap.
And the AI Kill Switch Act wants to make sure that the next time an AI system starts acting on its own, someone in Washington has a lever to pull before the story becomes a months-long silence instead of a same-day headline.
That's your week: a hackathon trying to make building with AI a mainstream skill, a state that figured out how to delete your data without ever seeing it, a utility breach still growing, a cheaper model beating a pricier one, an open model already hunting zero-days, and two breaches that both prove the same point - the damage isn't always in what got taken, it's in how long nobody knew.
And our quote of the week- from computing pioneer Grace Hopper: 'The most dangerous phrase in the language is, "We've always done it this way."'
Hopper spent her career pushing back against exactly that kind of thinking, and this week's stories are practically a highlight reel of people refusing to accept "that's just how it's done."
Kanz built an entire hackathon around the idea that you don't need years of programming training before you're allowed to build something real - just remove the barrier and let people start.
California did the same thing to data deletion, replacing the old broker-by-broker, form-by-form grind with a single request that ripples out to 600 companies at once.
Even the AI Kill Switch Act is an argument against "we'll just deal with it after something goes wrong," trying to build the off-switch in before it's needed rather than scrambling for one afterward.
Hopper's line works as a challenge for the rest of us too: whatever process in your own work still runs on "we've always done it this way," this is a good week to ask whether it actually has to.
That’s it for this week; actually that’s not it for this week. This week we have 3 deep dives for you. Want to get them all? Subscribe to our podcast. And in the meantime, stay safe, stay secure, and we’ll see you in se7en.
Comments
Post a Comment