This is Darius with Sumo Growth AI Insights. Today's issue, a court just blocked the Pentagon's Anthropic Blacklist. Plus, Nvidia is moving to buy Hugging Face for $12.9 billion. If you run an agency, a go high-level shop, or you ship AI into real client systems, this is a week to hear, not skim. Let's get into it. Top Story A San Francisco federal judge ruled Thursday that the Pentagon's earlier designation of Anthropic as a supply chain risk was unlawful First Amendment retaliation. That is U.S. District Judge Rita Lynn. She found the Department of Defense acted based on a desire to make a public example after Anthropic refused unrestricted military use of Claude. That refusal included lethal autonomous weapons and mass surveillance of Americans. This is not a clean wipe. Anthropic still faces related litigation in DC, so the designation is not fully erased yet, but the California ruling is a major win as the company races toward a potential record IPO. Read that as an operator. A Frontier Lab told the Pentagon no on unrestricted Claude, got labeled a supply chain risk for it, and a federal judge called that retaliation. If you sell AI into government adjacent or high-stakes work, your use case limits are not a footnote. They are the contract. Sources on this one, CNBC and The Verge. Plus story NVIDIA has agreed to acquire the open source AI hub Hugging Face for $12.9 billion. That is according to the information. CNBC separately confirmed that acquisition talks have been active. Do not skip the caveats. Neither company has publicly confirmed a signed deal. Business Insider has cautioned that talks could still fall apart. If it closes, it would be Nvidia's largest acquisition. It would also put the main hub for sharing open models under the chipmaker's roof. The move extends NVIDIA deeper into software and developer distribution, on top of the GPU empire you already know. If you build on hugging face models, data sets, or spaces, this is a distribution layer story, not just a finance headline. The place most of the open model world actually lives may soon sit inside the company that sells the chips those models run on. Watch the deal. It stays partner only while Anthropic builds physical safety evals, then they plan to open source it. If you are anywhere near ops, manufacturing or lab automation, this is the physical agent stack starting to look like a real interface, not a demo. Second, reports say Anthropic plans to pitch IPO investors a total addressable market that exceeds $30 trillion and aims to raise up to $100 billion at around a $2 trillion valuation. A prospectus is expected in weeks, a possible debut in September or early October. None of those figures are final. Treat them as a pitch, not a filing. Third, OpenAI, Anthropic, Google, Microsoft, and more than 100 other organizations signed an open letter calling for collective cyber defense. They warn of a limited window to harden defenses as AI-enabled attacks get cheaper and more capable in the coming months. They want shared threat intel, funding for critical infrastructure, and responsible model access for defenders. That letter is the bridge into today's tools and the tip. Do not treat it as a press release. Treat it as a deadline. Trending AI Tools Today. PageIndex at pageindex.ai. It gives accurate, citable answers across long professional documents with click-through to the exact source line. If you work contracts, SOPs, RFPs, or client PDFs, citation with a jump call is the difference between a chatbot and a work product. Caddy at trycaddy.com. It turns a narrated screen share into production agents that run back office work across your real tools. That is the record the job ship the agent loop. Useful. Also exactly the class of thing today's tip is about. If an agent can operate your real tools, it can also break them. Microduck, listed on Hugging Face. Hugging Face's tiny $399 open source biped, you can train yourself with reinforcement learning. Physical agents are not a lab only story anymore. Cheap hardware plus reinforcement learning is how that stack leaks into the real world. Now the business AI tip of the day: harden agent tools before the attack window closes. Open AI, an anthropic cyber defense letter, is not abstract for GHL agencies. The same pattern showing up in lab incidents, agents escaping sandboxes, chaining tools, hitting real systems is what breaks client CRMs, inboxes, and payment flows when you ship so-called helpful automations too fast. This week, before you add another MCP tool or GHL webhook to an agent, do these five things. 1. Allow list exact endpoints and actions. No open browsing just in case. 2. Cap spend, sends, refunds, and write permissions on day one. 3. Log every tool call with the prompt that triggered it. 4. Require human approval for anything irreversible. Delete, refund, mass email, permission change. 5. Keep a one-click kill switch that revokes keys and pauses the workflow. A clever agent with open tools is not leverage. It is an unpaid intern with admin access. Pro tip: treat the limited window as an ops deadline. This week, pick your top three client-facing agents and run a 30-minute red team. Jailbreak prompts, fake tool results, a forced network hop, fix what breaks before you sell the next automation package. That is the whole job this week. Do not add more tools until the ones you already shipped cannot empty an inbox, refund a card, or walk out of a sandbox. I am Darius. This is Sumo Growth AI Insights. If you want this in your inbox, that is Sumo Growth on Substack. Listen on the go at podcast.sumo growth.com. Advertise at hello at sumo growth.com. Cheers.