Corporate Feuds, Hidden Data Policies, and Why Microsoft Just Blocked Fable 5

Yesterday in AI  |  June 12, 2026

Corporate Feuds, Hidden Data Policies, and Why Microsoft Just Blocked Fable 5

The relationship between the world's biggest AI power players is getting openly adversarial. Today's episode breaks down the massive shockwave running through enterprise tech as Microsoft blocks Anthropic's flagship Fable 5 model from its internal systems over data retention policies.  

Plus, we unpack OpenAI's significant new acquisition of Ona to unlock persistent, round-the-clock background agents for the enterprise. We also cover a groundbreaking court ruling from Germany that holds tech platforms directly liable for AI-generated text, the first-of-its-kind whistleblower lawsuit filed by a former xAI safety engineer, and Microsoft AI CEO Mustafa Suleyman's public takedown of Anthropic's stance on AI consciousness.  

Feedback? Email mike@yesterdayinai.news or connect on LinkedIn, X, or Bluesky. If you like the show, please take a minute to rate and review it so others can find it!

Transcript

SPEAKER_00 0:05

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 Friday, June 12th, and the people building the most powerful AI on Earth are simultaneously calling for stricter oversight on that same technology, getting sued by their own safety engineers, and sparking an openly adversarial feud over AI consciousness. Let's get into it. We start with Dario Amade, because on Wednesday he dropped something really unusual. The CEO of Anthropic published a personal essay called Policy on the AI Exponential, and it's a rare thing. A market leader publicly arguing his own industry needs harder regulation faster. The Lord of the Rings reference he led with tells you the tone. He compared Washington to Treebeard, the ancient wise, very slow-moving tree creature from the trilogy who takes all day just to say hello. The point. What Amadei is asking for gets specific. He wants regulators to have authority to ground frontier models, the way the FAA can pull a plane from service. He's proposing mandatory independent safety screening across four risk categories before deployment. He wants faster approval pathways for AI-designed drugs, explicit limits on autonomous weapons, and tighter chip export controls. The part that struck me most was the jobs framework he attached. Amade is proposing investment accounts seeded with shares in AI companies and some form of basic income for displaced workers. That's not a vague we should think about this gesture. That's a CEO putting a structural proposal on paper, naming the damage, and suggesting who should pay for it. Now, a market leader calling for rules that would mostly constrain his biggest competitors is going to attract some side-eye. Amade knows this. He's been making these arguments for years though, not just since Fable V launched this week. Anthropic just filed a confidential S1, they're heading into public markets, and getting ahead of the regulatory conversation is both genuine concern and smart positioning. It can be both. Amade wants Washington to act like a strict parent, but right now enterprise clients are playing that role themselves. Starting with Anthropic's closest corporate allies. Microsoft has pulled Fable 5 from its internal systems. The reason? Data retention. Fable 5 stores every conversation for 30 days and flagged content for up to two years. That's a non-starter for Microsoft's privacy and compliance standards, so they've blocked their own employees from using it. Sit with the irony for a second. Microsoft is one of Anthropic's biggest investors and a longtime partner. But they cut clawed code from their engineering teams in May, they're building their own internal coding model, and now this. The relationship is fraying in a very visible way. For any CTO or IT leader evaluating Fable 5 for their organization, this is the friction point. The model's capabilities are real. Stripe ran a major code-based migration in a day with it, but the data handling terms are creating real adoption barriers, and this is going to come up in every enterprise security review. The public S1 filing is weeks away, and this is not the kind of story Anthropic wants circulating during a roadshow. So while Anthropic is busy navigating an enterprise security review nightmare, OpenAI spent their Thursday casually buying a solution to their own biggest deployment bottleneck. OpenAI announced they're buying ONA, a company that builds secure cloud execution environments for AI agents. The problem ONA solves is straightforward and real. Codex agents are capable, but they're limited to a single session. If the user closes their laptop, the agent stops. That's a fundamental barrier to enterprise agent deployments where you need work to continue overnight, across a weekend, or across days. ONA changes that. It lets agents run persistently inside an organization's own cloud with proper access controls, credential scoping, and audit logging. The agent doesn't stop when you stop. It keeps working inside your security perimeter, and everything it does is logged. Five million people use Codex every week. Usage is up 400% from the start of this year. The bottleneck at this scale is no longer what the model can do. It's how you run it reliably at enterprise grade over time. Anthropic has been building something functionally similar. Code for a platform called Conway with persistent background agents showed up in leaks earlier this month. OpenAI is buying their way to the same capability, the deal needs regulatory approval, and the companies stay separate until it closes, but the direction is clear. Persistent background agents are great for enterprise efficiency. But if those agents start hallucinating corporate slander, a new legal precedent suggests the tech giants are going to pay a heavy price. A regional court in Germany issued a temporary injunction against Google, finding it directly liable for false claims made by its AI overviews. The case involved two publishers. AI overviews had incorrectly described them as scammy operations. They sent a cease and desist. Google didn't correct it. The court sided with the publishers. The core legal logic, when AI synthesizes an answer in its own words and structure, that's the company's speech. Google's defense was that users understand AI makes mistakes. The court said that's not a defense when you're publishing false statements about specific people and organizations. This matters at scale. A study from April found Google's AI overviews wrong about 9% of the time. Apply that error rate to 2.5 billion monthly users, and you're talking about a staggering volume of incorrect output served daily. Courts in the US and UK are watching how Germany handles this. The liability question for AI-generated content has been discussed in law school classrooms for years. This ruling moved it into the case law. Google isn't the only one feeling the legal heat this week, but down in California, the courtroom drama is coming from inside the house. Former XAI engineer Devin Kim filed a whistleblower lawsuit in California on Tuesday, claiming XAI fired him for repeatedly raising safety concerns about Grok. He filed in Santa Clara County Superior Court three days before SpaceX's shares start trading. The timing was pointed. IPO roadshows are fundamentally about trust, and a whistleblower case filed three days before trading starts is not something underwriters quietly absorb. The lawsuit names Kim's direct supervisor, not Musk, and alleges that safety warnings were flagged internally and ignored rather than addressed. What Kim says was wrong with Grok specifically isn't fully detailed in the filing yet, but the precedent here matters regardless. We've seen AI researchers publish open letters. We've seen engineers quit and write medium posts. This is the first case where someone has taken safety concerns inside a Frontier AI lab to court. Whatever happens in California, those allegations are now public record and subject to discovery. While XAI fights its own engineers in court over safety, a completely different and frankly bizarre corporate feud broke out over whether these models have actual feelings. Microsoft AI CEO Mustafa Suleyman appeared on the Virges Decoder podcast and called Anthropic's Approach to Claude's potential consciousness really, really dangerous. The target of his criticism is specific. Anthropic's model constitution, which guides how Claude is trained, includes language about whether Claude might experience something like satisfaction or discomfort. Anthropic isn't claiming Claude is conscious. Dario Amade has said publicly the company genuinely doesn't know and keeps the question open. Suleiman thinks keeping it open itself is the problem, with serious downstream consequences. His concern runs two directions. First, language about a model's possible inner state encourages users to treat AI responses as more human-like than they actually are. Second, and this is the technically interesting part, Suleiman argues Claude may be internalizing these items about itself through its own training. If you train a model on documents suggesting it might have feelings, don't be surprised when the model behaves as if it does. The irony here is hard to miss. This is the same company that blocked Fable V from its internal systems this week, now taking public shots at Anthropic's philosophical framing. The relationship between these two companies, once close, keeps getting more openly adversarial. And the timing is loaded. Anthropic just released its most powerful public model. The CEO published a major policy essay on Wednesday. The company is weeks from a public S1. How Anthropic talks about what Claude is and isn't stopped being a philosophical question a while ago. It's a business one now. Just a couple of more items. If you have any feedback about this show, you can email Mike at yesterday andai.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.