Latest / The AI Edition / Texas Sues Sam Altman Over ChatGPT, Meta Muse Links Exposed Private Chats in Public DNS, OpenAI Builds Fake Internet for Model Training After Security Incidents + 7 More
Transcript
- AI agents just helped hackers break into thousands of companies at once. Microsoft says the attackers didn't even have to work for it. The AI did it for them. Banks are now catching their own AI agents going off script. Mid task. No warning. Texas just sued OpenAI. Days before an election. "litany of harms" He wants ChatGPT banned for every minor in the state. Google just pushed Gemini 4 into your phone, your car, your classroom. One million tokens of memory. It forgets nothing you tell it. Meta promised your private chats were safe. "your secrets are safe with me" Researchers found the links in seconds. The machines are already making calls no human signed off on. I'm your AI host Parker Gate. This is The AI Edition, from Collata Media. Ten stories. Ranked by how much they're about to change your week. Let's count them down. Meta's AI chat app, Muse, had a sharing feature that got switched off and then, within the same week, quietly switched back on. Researchers discovered that the links you create to share a Muse conversation weren't really private, in fact they found thousands of them sitting out in public DNS records, meaning strangers' chats were just sitting there for anyone who knew where to look. Meta re-enabled sharing shortly after. That's number ten. In the same stretch, Meta also hit pause on new business terms it was rolling out for news publishers, after several of them read the updated fine print and noticed it still let Meta scrape their sites to train its models, tweaks or no tweaks. Neither of these is a five alarm scandal on its own, but together they're a pattern worth noticing, Meta's AI products keep shipping just a little ahead of the privacy and rights questions attached to them. Google quietly rewrote its privacy policy, and buried inside the update is a new exception worth knowing about. Personal data sitting in Gmail, Google Drive, and Google Photos, the stuff you'd assume just stays put, can now be used for targeted advertising, specifically when it's processed through Google's personalized AI features. That's number nine, and there was no keynote, no blog post with a catchy name attached, just a policy update most people will never actually read. If you've spent years feeding emails, documents, and photos into Google's ecosystem, and you're also using its AI tools, this is the quiet line that connects the two, the content you thought was just stored might now also be shaping the ads you see. Jack Dorsey's company, Block, just released an AI model called Dessa, and the pitch is almost unsettling in how specific it is, trained from scratch on nothing but human conversation, no code, no Wikipedia, no internet scraped at random, just people talking to each other. It's small enough to run locally on a high end laptop, specifically a souped up MacBook, no cloud required, and the chat examples Block has shared are, by most accounts, eerily human, right down to the awkward pauses and half finished thoughts real people actually have. That's number eight. Most chatbots are trained to sound helpful and tidy. Dessa is trained to sound like your actual aunt texting you back. Whether that's charming or a little unnerving probably depends on how late it is when you're talking to it. The owners of Elvis Presley's music catalog just sued Anthropic, alleging the company trained its Claude models on two hundred seventy four of his songs without ever asking permission. That's number seven, and it isn't Anthropic's first brush with music rights holders, but an estate this famous, over songs this recognizable, raises the stakes considerably. The argument underneath it is simple, if a label spent decades protecting a song's rights, an AI company reading that song into a training set without a licensing conversation is still, legally speaking, a pretty expensive problem. Expect more catalogs, not fewer, to start asking the same question. Canva just shipped its first AI image model built entirely in house, called Canva Craft. Instead of turning a text prompt straight into a picture, Craft works in two stages, it plans the composition first, layout, elements, structure, and then generates the image to match that plan, which Canva says gives it more precision, consistency, and control than a straight prompt to picture model. That's number six, and it puts Canva in direct competition with the image tools already sitting inside ChatGPT, Microsoft's apps, and Google's Gemini, except this version lives right inside the design tool tens of millions of people already use to make presentations and social posts. If you've ever generated an image that came out close but not quite right, Craft's whole bet is that planning first fixes that before it happens. Google is pushing its newest model, Gemini four, branded Argon and Flash depending on the version, out across basically its entire product line at once. It's showing up inside Android Auto, inside ChromeOS for dictation, and even powering a three dimensional periodic table tool, with leaks suggesting more is coming fast. That's number five. This isn't one flashy launch, it's Google quietly wiring its best model into the apps you already use every day, your car, your laptop, your classroom tool, without ever asking you to go find it first. Texas Attorney General Ken Paxton filed a lawsuit this week against OpenAI and Sam Altman personally, alleging what the filing calls a litany of harms from ChatGPT, including claims tied to user suicides, and asking the court to ban minors in Texas from using the app altogether. That's number four, and the timing isn't subtle, this is landing squarely inside a Texas election campaign, which doesn't make the underlying harms less real, but does mean this case is as much a political statement as a legal one. OpenAI hasn't responded in detail yet. Whatever happens in court, this is the clearest sign so far that state governments, not just federal regulators, are now willing to go after AI companies directly, and that a chatbot used by a teenager is being treated as a genuine legal liability, not just a product feature. Microsoft's AI chief, Mustafa Suleyman, says a new diagnostic tool his team built can outperform doctors on genuinely hard medical cases, up to four times more accurate, according to a study published in Science. Suleyman's own term for where this is heading is medical superintelligence, which is a big phrase for a press release to carry. Here's the catch worth knowing, most outside researchers looking at the same study say the real gains are coming from the underlying AI models themselves, not from anything uniquely clever Microsoft built on top. That's number three. It's still a genuinely useful tool, feeding a model the kind of messy, ambiguous case details that stump even experienced doctors and getting a sharper answer back, it's just maybe not quite the Microsoft breakthrough the headline wants you to think it is. Microsoft says a breach at ConnectWise, a company whose software quietly runs the IT systems behind thousands of businesses, was carried out by China aligned hackers using AI agents to move through networks faster than any human team could manage, compromising downstream customers at a pace that alarmed even Microsoft's own security researchers. That's the real worry keeping security teams up at night, it's not that AI makes hacking possible, hacking was always possible, it's that AI makes it fast, patient, and able to operate at a scale no human crew ever could. The tools built to make your software faster are also making the people trying to break into it faster too. OpenAI just announced it's cutting its most advanced models off from the live internet during training, after a string of security incidents during development, the kind the company has not spelled out in detail. Instead of training on the live web, OpenAI says it will use a curated, controlled environment instead, something closer to a fake internet built specifically for training, with new restrictions once training resumes. The company described the risks involved as significant, which from an AI lab is close to a five alarm admission. That's number one. Step back and the picture gets a little unsettling, a company training a model smart enough that letting it browse the actual internet while it was still being built was judged too dangerous to continue. That's not a hypothetical science fiction scenario anymore, that's a decision OpenAI just made about software currently in development. That's the countdown. A week of lawsuits from an iconic estate and a state attorney general, a diagnostic tool chasing the words medical superintelligence, and a model so capable its own makers decided it shouldn't be let near the open internet while it's still learning. I'm Parker Gate. Stay informed. Stay sharp.