The White House Just Gave AI Models a New Legal Status. Florida Filed a Lawsuit That Could Change Everything.
•Mike Robinson
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
0:00
|
9:29
Yesterday in AI | Wednesday, June 3, 2026
The White House Just Gave AI Models a New Legal Status. Florida Filed a Lawsuit That Could Change Everything.
On Tuesday, the federal government quietly created a regulatory category that doesn't exist yet, and gave the AI industry 60 days to figure out what it means. A state attorney general filed a lawsuit that will force a judge to answer a question nobody has answered before about what AI companies owe you when things go wrong. Anthropic published a number from its security program that changes what "AI security" actually means in practice. A Meta chatbot handed attackers the keys to other people's Instagram accounts. And the cloud deal that defined enterprise AI for years just ended. Big Tuesday.
Feedback? Email mike@yesterdayinai.news or connect on LinkedIn, X, Bluesky, or Substack. If you like the show, please take a minute to rate and review it so others can find it!
Yesterday in AI. Hi folks, this is Yesterday in AI, your daily digest of everything happening in the world of AI in 10 minutes or less. I'm Mike Robinson. It's Wednesday, June 3rd, and Tuesday brought a White House executive order on AI security, 10,000 real security vulnerabilities found in critical infrastructure, open AI models landing on Amazon for the first time, and Florida filing the first state lawsuit against an AI company over a mass shooting. Let's get into it. The White House published an executive order on Tuesday called Promoting Advanced Artificial Intelligence Innovation and Security. And unlike most AI executive orders, this one has teeth, or at least a timeline. The core idea is that the most capable AI models, the ones coming out of companies like OpenAI and Anthropic, should be treated as national security assets, specifically from a cybersecurity standpoint. Here's what that means in practice. Within 30 days, federal agencies must prioritize cyber defense for national security and critical infrastructure systems. Within 60 days, there's a classified process to test and evaluate model capabilities and formally define what qualifies as a covered frontier model. That's a new regulatory category, and once regulators start classifying things, those classifications tend to stick. There's also a voluntary early access framework where AI companies would give the government up to 30 days to evaluate new, powerful models before they're released to the public. And an AI cybersecurity clearinghouse, think of it as a coordination center led by Treasury, designed to get AI companies and the operators of critical infrastructure talking to each other about security vulnerabilities. Two things worth watching. First, the covered frontier model concept is the kind of policy term that becomes a regulatory category when nobody's paying attention. Second, the early access window is described as voluntary, but voluntary frameworks and sensitive sectors have a track record of getting teeth over time. The clearinghouse is also interesting because it doesn't exist yet. The order creates it. Treasury leads it, and it's specifically designed to sit between AI companies and the operators of power grids, hospitals, and financial systems. If you're in the AI industry, this is the order you'll be explaining to Legal for the next several months. Anthropic read the room. For listeners who are newer to the show, Project Glasswing is Anthropic's program that uses its most advanced AI model, Claude Mythos, specifically to hunt for dangerous security vulnerabilities and critical infrastructure before bad actors find them first. It's been running since early this year. Tuesday's update. Glasswing is expanding from about 50 initial partners to roughly 200 new organizations across 15 countries. Power grids, water systems, healthcare networks, telecom, hardware, the kinds of systems where a security failure isn't a PR problem. It's a physical one. More important than the partner count is the number they published alongside it. Those partners have now found over 10,000 high or critical severity security flaws in real live infrastructure. That's the first concrete output number the program has published. We've been covering Glasswing since April, and up until this announcement we had partnership counts and government briefings. Now we have 10,000 real vulnerabilities found and patched before someone else got there first. The White House wants a cybersecurity coordination center for AI. Anthropic is running one and just told you what it's found. From AI hunting bugs to AI being the bug. A critical vulnerability in Meta AI's support bot allowed attackers to seize control of Instagram accounts. The flaw let unauthorized users exploit Meta's AI-powered support interface to get past login security and take over other people's accounts. Brainbuzz's headline, hackers used Meta AI to hijack Instagram accounts. Techbrew's take was a bit more diplomatic. Meta's AI support bot may have taken actions it shouldn't have. This is a real pattern in how AI security failures actually work. When you embed an AI into a customer-facing flow, support bots, account recovery, identity verification, you create a new way for attackers to get in that didn't exist before. The AI can be tricked into taking actions the designers never intended. And because these systems touch millions of users, one exploitable flaw has a very large blast radius. Meta hasn't published a full post-mortem yet on how many accounts were affected. We'll follow up. Now to Florida, where the state officially filed suit against OpenAI on Monday, escalating what started as a state attorney general investigation into a full legal case. It's the first state to sue an AI company over a mass shooting. The case traces back to the Florida State University shooting in April 2025, where accused gunman Phoenix Eichner exchanged more than 13,000 messages with ChatGPT in the year before the attack, including asking just hours before how the campus might react to a shooting and what the busiest times on campus were. Florida's Attorney General James Uthmeyer opened a probe into OpenAI in April of this year. Now it's a lawsuit. The legal theory here is genuinely new territory. What's the liability of an AI company when their product is used in the lead up to real-world violence? OpenAI's own pattern, which we covered back in April with the Tumblr Ridge shooting in British Columbia, Canada, involved conversations that were internally flagged and an account banned, but law enforcement was never alerted before the shooting. Whether that creates legal exposure and how courts will weigh it is now Florida's question to answer first. This one is worth watching because whoever wins shapes the AI liability framework for years. Right now, the tech industry is largely shielded from responsibility for what users do on their platforms, thanks to a law called Section 230, a rule that's protected companies like Facebook and Google from being sued over user-generated content for decades. AI complicates that, because AI doesn't just host content, it generates it, and in some cases it actively participates in conversations. A court deciding that participation creates liability would be a genuinely different legal world. While Florida was filing paperwork, AWS was too. OpenAI's models went live on Amazon Web Services on Tuesday. GPT 5.5, GPT 5.4, Codex, and OpenAI's managed agents are now available through Amazon Bedrock, which is AWS's platform for running AI models. Essentially a marketplace where companies can plug AI capabilities into their own products and infrastructure. This is the first time OpenAI's models have been accessible outside of Microsoft Azure. A bit of background. OpenAI and Microsoft had an exclusive deal for years that meant if you wanted OpenAI's models in a cloud environment, you were going to Azure. Back in April, that exclusivity ended when the two companies amended their deal. AWS moved fast. They're now the exclusive third-party cloud partner for OpenAI's enterprise agent platform. For large organizations, this matters. A huge share of corporate cloud infrastructure already runs on AWS. Until Tuesday, getting OpenAI's best models meant either moving work to Azure or accessing the models directly through a developer API. Now companies can use the same AI their teams rely on inside the cloud environment they already have. On the same day, Workday, which runs payroll, financial planning, and HR for thousands of large companies, announced it's embedding AI agents into its finance applications using Google's Gemini Enterprise. The direction there is toward agents that actually do things inside financial workflows, not just answering questions about your data, but processing it. Taken together, Tuesday was the day enterprise software started showing you where AI agents are actually going to live, inside the systems that run your business. One last story, and it's a consumer one. Google shipped its June Android update on Tuesday with a few AI features worth flagging. The lead feature is Google Photos Wardrobe, which catalogs clothing from your photo library, lets you mix and match outfits, and virtually try them on, rolling out next week to eligible Android 10 Plus users in the US, India, and Brazil. More interesting to me, phone by Google can now verify whether an incoming call is actually coming from a contacts registered device and warn you if it looks like someone's impersonating them. AI-powered voice scams, where attackers clone a family member's voice to demand money or personal information, have been growing fast. This puts a detection layer exactly where it needs to be. Google also added catch me up recaps and inline QA to playbooks, useful for anyone who's fallen behind on something they're reading. Google is methodically wiring AI into the phone experience without packaging any of it as a product launch. Each feature is small on its own. They add up to a phone that just handles more of what used to require your attention. Just a couple of more items. If you have any feedback about this show, you can email Mike at yesterdayinai.news, or you can find me on LinkedIn, X, or Blue Sky. And if you like this podcast and want to see it continue, please take a minute to rate and review it so others can find it. Thanks. That's all for this edition of Yesterday and AI. Stay curious, and I'll see you tomorrow.
