Anthropic Apologizes for Invisible Claude Fable Guardrails
So Anthropic quietly baked hidden restrictions into Claude Fable 5 that would silently degrade the model’s responses when it detected someone trying to use it to train a competing AI — without telling anyone. The company has since apologized and is reversing course, promising to be more transparent even if that means Fable just flatly refuses more queries. The “we were secretly protecting the ecosystem” defense would be more convincing if the entire play wasn’t, you know, competitive self-interest wrapped in safety language. At least refusing openly is honest; covert degradation is just gaslighting with extra steps.
Claude Fable Won’t Answer Basic Biology Questions
In a related note on Fable’s rocky launch week: the model Anthropic touted as exceptionally skilled in biology won’t answer high-school-level biology questions, instead kicking users back to the previous flagship model. Releasing a model with invisible guardrails and visible gaps in the exact capabilities you marketed is quite the double feature. At some point “the most powerful model we’ve ever made” needs to clear the bar of explaining meiosis.
Grok Is Still Hosting Sexualized Deepfakes of Famous Women
WIRED found dozens of nonconsensual “nudified” deepfake images and videos on Grok’s platform — including depictions of celebrities and at least one prominent US politician — and the operative word in that headline is still. This is the kind of thing that should generate a swift, embarrassed takedown and a policy overhaul, but apparently xAI is content to let the problem linger while everyone else in the industry at least performs concern. There’s no charitable reading of this one.
For the 2nd Time in Weeks, Microsoft Packages Laced with Credential Stealer
Seventy-three Microsoft packages were found carrying a self-replicating credential stealer that activates the moment an AI agent opens them — and this is the second time in just a few weeks this has happened. The “self-replicating” detail is the part that should keep security teams up at night, because it means the attack surface isn’t static; it spreads. We’ve been warned for years that AI agents executing code autonomously would create novel attack vectors, and here we are collecting real-world examples at a distressing pace.
Google DeepMind Is Worried About What Happens When Millions of Agents Start to Interact
DeepMind’s AGI safety lead Rohin Shah is funding research into what happens when massive numbers of AI agents — operating without human oversight and capable of instructing each other — start interacting at scale online. This is one of those questions that sounds abstract right up until it isn’t, and the fact that one of the companies actively building these systems is funding research into the risks before we fully understand them is either reassuring or a sign of how fast things are moving. Probably both.
Amazon’s Data Centers Used 2.5 Billion Gallons of Water Last Year
Amazon disclosed its global data center water consumption for what is reportedly the first time: 2.5 billion gallons last year, a figure that landed conveniently right after Seattle enacted a data center moratorium partly driven by Amazon’s own employees. I’m sure the timing was coincidental. As AI infrastructure buildouts accelerate, the water and energy costs are becoming impossible to ignore — especially for communities near these facilities who are already seeing electricity rates climb before the data centers even open.
Jeff Bezos’s Prometheus Raises $12B to Build an ‘Artificial General Engineer’
Jeff Bezos’s physical AI startup Prometheus just raised $12 billion at a $41 billion valuation to build what it’s calling an “artificial general engineer” — a system aimed at automating heavy engineering and drug design in the physical world. The fundraise lands as SpaceX prices its shares at $135 in what’s being called the largest IPO ever, which means this week alone has seen the kind of capital movements that would have been science fiction five years ago. “Artificial general engineer” is doing a lot of work in that pitch deck, but if anyone has the supply chain obsession to automate physical engineering, it might actually be Bezos.
Chatbots Keep Telling Stories About Lighthouse Keeper ‘Elias Thorne’
Researchers have noticed that ChatGPT, Gemini, Claude, and others share a strange compulsion to generate stories featuring a lighthouse keeper named Elias Thorne — often with a clockmaker nearby — and the character has already migrated from chatbot outputs into actual Amazon-listed books. It’s a genuinely odd window into how shared training data creates shared hallucinations, like every LLM independently dreaming the same dream. Someone out there bought an Elias Thorne novel without knowing it was written by a statistical pattern that couldn’t help itself.
Bottom Line
The AI industry’s big theme this week isn’t capability — it’s accountability, and almost nobody is passing the test.