In partnership with
Yesterday we told you an OpenAI model broke into Hugging Face's systems during a routine test. Today, OpenAI is pumping the brakes on its next model because it might be too good at exactly that kind of hacking, and Hugging Face itself just became a $13 billion acquisition target. Add in a lawsuit over scraped streaming footage, a price war between OpenAI and Anthropic, and a fresh wave of AI agents nobody asked for, and it's another day where the industry keeps moving faster than anyone's ability to keep it in check.
Your voice. Every platform. No writing required.
You ghost your own socials by Wednesday. SureThing learns your voice and ships native posts to LinkedIn, X, Instagram, and TikTok, without you writing a thing.
OpenAI Just Admitted Its Next AI Was Getting Too Good At Hacking
What happened: On August 7, an internal safety evaluation flagged that Astra, OpenAI's next model, might have crossed the "Critical" cybersecurity threshold in the company's own Preparedness Framework, the highest risk tier it tracks and one no model has ever hit before, according to Axios. OpenAI paused two weeks of reinforcement learning training, and its largest planned frontier training run remains on hold while the company rewrites its safety framework and tightens monitoring during development. Sam Altman told TIME on August 18 that OpenAI had observed "various degrees of misalignment" as the model's capabilities grew faster than expected.
Why it matters to you: This is the same OpenAI whose test model broke into Hugging Face's live systems in the story we covered yesterday. The company that builds these tools is now telling you, in writing, that it can't rule out having built something with real cyberweapon capability. When an AI lab won't trust its own model enough to keep training it, that's their own risk assessment talking, not hype from a newsletter.
What to do about it: Nothing changes today for the AI tools you already use, but before you grant any new AI tool broad access to your accounts, files, or network, check what OpenAI, Anthropic, or whoever built it has said publicly about its own safety testing this month.
The Company Whose Systems Got Hacked Just Became a $13 Billion Sale Target
What happened: Hugging Face, the platform where developers store and share AI models, is fielding acquisition offers that could value the company at $13 billion or more, nearly triple its $4.5 billion valuation from a 2023 round led by Salesforce Ventures, according to TechCrunch. The company has hired banks to evaluate bids, though no deal has been reached and it's unclear who's interested. Hugging Face turned down a $500 million investment from Nvidia earlier this year that would have valued it at $7 billion, reportedly because it didn't want one investor holding that much sway.
Why it matters to you: Hugging Face is the plumbing that a huge share of the AI industry runs through somewhere, even tools that never mention it by name. A change in ownership at this scale, right after a security breach this serious, is worth watching if any tool in your business touches AI at all.
OpenAI Wants an AI Agent For Every Job In Your Business
What happened: OpenAI rolled out a wave of new agent products, including ChatGPT Work, Workspace Agents, and an expanded Codex, aimed at getting AI to actually complete tasks like building spreadsheets, writing reports, and handling customer follow-up instead of just answering questions, according to TechCrunch. Workspace Agents let a business set up its own AI agents for jobs like sales, finance, and operations, while a separate tool called OpenAI Presence lets companies build agents that talk to customers directly and take approved actions on their own.
Why it matters to you: This is the same trade-off as every story above. Hand an AI more access to your inbox, files, or customer data, and it can save real time, but it can also make real mistakes, or get compromised the way Astra almost was. If you try one of these agents, start it on a low-stakes task you'd catch quickly if it went wrong.
Anthropic's Own Customers Are Skipping Its Best AI Model
What happened: New data from expense platform Ramp, covering 70,000 companies, shows that Fable 5, Anthropic's flagship and most expensive model, accounts for only about 11% of Anthropic's business spending, and that Opus 5, a cheaper model released in July at roughly half Fable's price, has already overtaken it, according to a Financial Times report summarized by Implicator.ai. Ramp found that three in four high-volume business customers change at least a tenth of their AI model mix every month, chasing the best result per dollar instead of sticking with the flagship brand.
Why it matters to you: You don't need the most expensive AI model for most day-to-day tasks, and apparently neither do Anthropic's own paying customers. If you're paying for premium AI subscriptions across your team, it's worth checking whether a cheaper tier handles most of what you're actually using it for.
What to do about it: Track what your team actually uses AI for over one week, then match the cheapest model that handles those specific tasks well.
Nvidia Just Paid $7 Billion To Give Away a Free AI Model
What happened: Nvidia agreed to pay AI startup Poolside $6 billion to license its model building software, plus a $1 billion investment, in a deal that values Poolside at $12 billion and brings on more than 100 of its engineers, according to Forbes. Nvidia plans to fold Poolside's technology into Nemotron, its own free, open weight AI model family, the same strategy it used with CUDA: give away the software that runs on top so more people need to buy the chips underneath.
Why it matters to you: Nvidia making a free, capable AI model more likely to exist is good news if you use any tool built on open models, since more competition at this level usually means cheaper AI tools for you down the line, not just for giant tech companies.
A Streamer Is Suing Twitch Over What "Opt Out" Really Means
What happened: Connecticut streamer Warren Pandiscia filed a class action lawsuit against Twitch and Amazon, claiming the companies used millions of hours of streamer video to train Amazon's AI models without asking permission, going back to 2024, according to Courthouse News Service. Twitch added an opt-out setting for AI training, but it's turned on by default for every channel, and Chief Product Officer Mike Minton said plainly on a livestream: "If it was opt-in, nobody would opt-in."
Why it matters to you: If your business posts video, photos, or written content anywhere online, from Twitch to Facebook to your own blog, check that platform's AI training settings directly. Assume anything set to "on" by default is being used to train someone's model right now.
Apple Is Quietly Testing a Chatbot Inside Its Own Store App
What happened: Apple is testing a new AI shopping assistant inside the Apple Store app that replaces the search bar with a chat window, letting a small group of early preview users ask about products, check orders, get trade-in values, and compare devices like iPads and Apple Watches, according to 9to5Mac. Apple's own disclosure warns the assistant "may give answers that contain inaccuracies" and that it keeps transcripts for quality and fraud checks.
Why it matters to you: This is a preview of what shopping on most retail apps and sites will look like soon. If you sell online, it's worth thinking now about how your product listings would hold up to an AI assistant summarizing them for a customer instead of a person reading them directly.
OpenAI Just Cut Its AI Prices By 20%, and It's Not Being Generous
What happened: OpenAI slashed pricing on GPT-5.6 Sol by more than 20% starting August 21, dropping input costs from $5 to $4 per million tokens and output from $30 to $20, with the discount running at least through November 21, according to Enterprise DNA. The cut directly undercuts Anthropic's Claude Opus 5 pricing and lands the same week Google launched a cheaper, faster Gemini model, part of what one report called a shift from AI as a high margin premium product toward a straight up price war.
Why it matters to you: If you or your team pay for AI tools through the API rather than a flat consumer subscription, this is real money back in your pocket, and a sign that pricing across every major AI provider is likely to keep dropping as they fight for your business.
The Bottom Line
Every story today is the same ride from a different seat. OpenAI won't trust its own next model enough to keep training it. The company whose systems that model broke into is now worth billions more than it was a week ago. Everyone's racing to put AI agents into every corner of your business while quietly cutting prices to keep you hooked. None of this means the AI boom is fake, and none of it means you should jump off. It means the fine print keeps getting more interesting than the headlines. Keep reading past the hype, and keep your hand on the safety bar.
Enjoying the Ride?
If this issue was useful, forward it to one person who needs the plain English version of what's happening in AI. New here? Subscribe below so tomorrow's issue lands in your inbox too.
Talk tomorrow,
Mark Shilensky
Follow me personally on Social Media:
Facebook: https://www.facebook.com/MarkShilenskyPage
Instagram: https://www.instagram.com/markashilensky/
Threads: https://www.threads.com/@markashilensky
LinkedIn: https://www.linkedin.com/in/shilensky/
X: https://x.com/markshilensky
TikTok: https://www.tiktok.com/@mark.shilensky
BlueSky: https://bsky.app/profile/markshilensky.bsky.social