Pentagon’s AI Ultimatum, Anthropic’s Cowork Expansion & Video-Trained AI Agents

Pentagon Hits Anthropic With Ultimatum Over Guardrails 🛡️

The US military is threatening to draft Silicon Valley. The ideological war between the Department of Defense and Anthropic has reached a breaking point. According to reports from Axios, Defense Secretary Pete Hegseth has delivered a strict Friday ultimatum to Anthropic CEO Dario Amodei: remove Claude’s military safeguards, or face devastating consequences.

Here are the details of the unprecedented standoff:

• The Red Lines: Anthropic has consistently refused to allow its models to be used for two specific purposes: mass surveillance of American citizens and fully autonomous weapons (kinetic strikes without a human in the loop).
• The Ultimatum: Hegseth reportedly gave Amodei three choices. The first is to agree to “all lawful purposes” and drop the guardrails. The second is to have their $200 million contract killed and be labeled a “supply chain risk” (barring all defense contractors from using Claude).
• The Nuclear Option: The third and most extreme threat is forced compliance via the Defense Production Act, a Cold War-era law that allows the government to mandate private companies to prioritize national defense orders.
• The Replacements: While Anthropic holds the line, the Pentagon is actively fast-tracking classified access for OpenAI and Google, having already signed an agreement with Elon Musk’s xAI (Grok), which agreed to the unrestricted terms.

Why it matters: This is a terrifying precedent. The entire premise of “AI Safety” relies on the idea that the companies building these “god-like” systems will responsibly restrict their most dangerous capabilities. If the US government uses legal threats to forcefully strip away those safety guardrails, the line between an intelligent assistant and an autonomous weapon system completely evaporates.

UrviumAI Take: The invocation of the Defense Production Act changes everything. Pay close attention to how other tech CEOs react to this. If the government can legally force a private AI lab to build targeting software against its will, the concept of “Constitutional AI” is dead. The military, not the developers, will dictate the alignment of artificial general intelligence.

Anthropic’s Cowork Platform Escalates Agent Wars 🏢

Claude is applying for a job in every department of your company. Escalating the enterprise agent war against OpenAI’s “Frontier” platform, Anthropic has rolled out a massive update to its Cowork ecosystem, turning Claude into a fleet of highly specialized digital employees.

Here is how Cowork is infiltrating the enterprise stack:

• Department-Specific Agents: Anthropic launched pre-built AI agents tailored for 10 specific departments right out of the gate. This includes specialized bots for HR, software engineering, banking, equity research, and wealth management.+1
• The Connectors: Agents are useless if they can’t touch your data. Cowork now features direct integrations with Google Workspace, DocuSign, FactSet, and Harvey, allowing Claude to pull context and execute actions across multiple SaaS platforms.
• Private App Stores: Companies can now build “private agent stores,” curating custom bots for specific teams while maintaining strict admin controls over who can access what data.
• Cross-App Agility: A new research preview demonstrated Claude’s ability to hop between Microsoft applications autonomously—crunching complex financial data in Excel and immediately building a fully formatted pitch deck in PowerPoint.

Why it matters: The narrative has shifted from “AI that helps you type” to “AI that does your job.” By building out industry-specific agents and connecting them to the enterprise plumbing (DocuSign, Salesforce, Workspace), Anthropic is ensuring that Claude becomes the invisible connective tissue of the modern knowledge economy.

UrviumAI Take: The Cross-App Agility (Excel to PowerPoint) is the holy grail for corporate productivity. If you are a knowledge worker, start transitioning from “creator” to “editor.” Tools like Cowork mean you will no longer build presentations from scratch; your value will come from reviewing, verifying, and refining the automated outputs generated by your department’s specific agent.

New AI Learns Any Computer Task by Watching Video 👁️💻

AI is no longer just reading the internet; it’s watching how we work. A startup called Standard Intelligence has introduced FDM-1, a groundbreaking “computer action” model that completely changes how AI learns to operate software.

Here is how this visual approach to AI works:

• Video Training: Instead of reading code or text, FDM-1 was trained on 11 million hours of screen footage—a dataset roughly 550,000 times larger than previous open datasets. The AI essentially reverse-engineers the exact mouse clicks, drags, and keystrokes that produced each frame of the video.
• Massive Context: The model is capable of watching and tracking nearly two hours of continuous screen activity at once, retaining 50 times the visual context of existing models.
• Real-World Execution: Because it learns visually, FDM-1 isn’t limited to APIs or text-heavy software. Demos showed the model successfully building 3D gears in Blender (CAD), finding software bugs, and even using arrow keys to drive a real car in San Francisco via live data feeds.

Why it matters: This is a massive leap for agentic AI. Language models are great at text, but the modern computer interface is visual and spatial. By training a model directly on video footage of humans working, Standard Intelligence has unlocked the ability to automate highly complex, proprietary desktop software that traditional AI struggles to navigate.

UrviumAI Take: Training on screen recordings is the ultimate shortcut. If you have highly specialized software at your job (like legacy inventory systems or proprietary design tools), the days of “AI can’t use this” are over. Soon, you will just record yourself doing a task once, feed the video to a model like FDM-1, and it will perfectly replicate your workflow forever.

Last AI News: Anthropic Exposes Chinese AI Theft, OpenAI’s Consulting Alliance & OpenClaw Goes Rogue

Other AI News Today:

• Google Labs acquired AI music creation platform ProducerAI, integrating it with DeepMind’s Lyria 3 model to give artists granular control over audio generation and custom virtual instruments.
• Anthropic introduced Remote Control for Claude Code, allowing developers to start a coding task in their terminal and manage the session from their phone or browser.
• Meta announced a multi-year partnership with AMD to deploy up to 6 gigawatts of AMD Instinct GPUs, a massive strategic move to diversify its AI infrastructure away from Nvidia.
• OpenAI hired Arvind KC, a veteran of Roblox, Google, and Palantir, as its new Chief People Officer to manage scaling and redefine work as AI agents take over traditional corporate roles.
• Inception Labs launched Mercury 2, a diffusion-based language model that generates over 1,000 tokens per second, fundamentally changing the speed and cost equations for production AI.

Jigar Chaudhary

Jigar Chaudhary is the Editor-in-Chief at UrviumAI, where he oversees coverage of artificial intelligence news, tools, and in-depth studies. With over 5 years of experience analyzing AI and robotics, he focuses on maintaining high editorial standards, accurate reporting, and clear explanations to help readers understand how AI is shaping the future.