Three separate AI agent tools became available this week that run in the background while you sleep.

Yet, there’s a good chance you missed most of that because the U.S. government and top AI labs were either straight up fighting or making massive deals with far-reaching implications. 

OpenAI locked in the four biggest consulting firms on the planet to deploy AI that does exactly what those firms charge for. 

Yeah. A lot happened.

Here's your complete breakdown of the AI news that matters.

OpenAI CEO Sam Altman announced this past week that OpenAI reached an agreement with the Department of War, giving the Pentagon access to its AI models on classified networks.

This came directly after Anthropic failed to land a similar deal, refusing to allow its technology for fully autonomous weapons or mass domestic surveillance.

More than 60 OpenAI employees and 300 Google employees backed Anthropic's position in an open letter.

Altman confirmed the agreement includes human oversight for any use of force, protections against domestic surveillance, and a termination clause if the contract is breached. OpenAI engineers will deploy directly alongside Pentagon staff.

Altman publicly called for de-escalation, asking that similar terms be extended to all AI companies industry-wide. All of this broke right before news of US and Israeli military action against Iran.

What it means: This deal almost certainly took months to negotiate. It was not a reaction to Anthropic's standoff. Altman's de-escalation call suggests he understands the precedent being set. With AI leaders heading to the White House this week, Anthropic's conspicuous absence from the guest list is doing a lot of talking.

Welp.

President Trump ordered every US federal agency this past week to immediately stop using Anthropic's technology, giving a six-month window to phase it out entirely.

Defense Secretary Pete Hegseth designated Anthropic a supply chain risk to national security after the company refused to allow its models for fully autonomous weapons or mass domestic surveillance. No US company has apparently ever received that designation before.

Anthropic, which signed a $200,000,000 Pentagon contract last July, vowed to challenge the move in court. The company found out the same way everyone else did: through a Truth Social post.

Claude shot to number one on the US Apple App Store in the immediate aftermath. If fully enforced, companies with both federal contracts and Anthropic models will face a very real choice.

What it means: This is uncharted territory. Anthropic has a legitimate legal fight ahead. But any company holding federal contracts and currently running on Claude needs to be asking hard questions right now, not next quarter. The downstream implications for hundreds of businesses could be massive.

Google launched Nano Banana 2 this past week and claimed the number one spot on Arena, the AI image generation benchmarks that OpenAI had previously held across the board.

Built on the Gemini 3.1 flash backbone, it delivers better prompt adherence, accurate text rendering, and full creative control at half the cost of Nano Banana Pro.

Already rolling out across 141 countries in eight languages. Gemini Pro and Ultra subscribers get direct access to Nano Banana 2 inside Google Gemini, with resolutions from 512 pixels up to 4K and two quality settings.

The most important new capability: a built-in image search tool that grounds generation in real search results, unlocking enterprise workflows like storyboarding, product photography, and infographics. Subject consistency holds across up to five characters and 14 reference objects.

What it means: AI images finally stopped looking AI-generated. At half the cost of its predecessor, Nano Banana two makes image generation practical well outside creative and marketing roles. Better text rendering plus search grounding changes what these tools can actually do for your business. This one applies more broadly than most people realize.

OpenAI announced multiyear partnerships with Accenture, Boston Consulting Group, Capgemini, and McKinsey under a new initiative called Frontier Alliances.

The mission: deploy OpenAI's enterprise AI platform, Frontier, across their global client bases at scale. Frontier lets businesses build, deploy, and manage AI agents across company systems, and enterprise clients already account for roughly 40% of OpenAI's revenue.

These firms will build certified teams, work directly alongside OpenAI's forward-deployed engineers, and receive early access to OpenAI's product roadmap. Chief Revenue Officer Denise Dresser said the partnerships leverage the firms' deep client relationships to meet surging demand.

BCG and McKinsey handle strategy. Accenture and Capgemini handle technical integration.

What it means: Let's be honest. These are the same firms charging premium rates for research and deck production that today's agentic AI handles by default. Watch how their pricing models evolve. The billable hour is being renegotiated in real time, and if you engage any of these firms, this development changes the conversation you should be having with them.

More AI drama? Yup.

Anthropic publicly accused Chinese companies DeepSeek, Moonshot AI, and MiniMax of running coordinated, industrial-scale campaigns to copy Claude's most advanced capabilities.

Over 16 million interactions. More than 24,000 fraudulent accounts. One proxy network managing over 20,000 fake accounts simultaneously to avoid detection.

The technique is called distillation: training your own model on the outputs of a stronger one.

OpenAI sent a memo to Congress about similar attacks in the same two-week window. Google identified over 100,000 distillation prompts targeting Gemini.

Anthropic's share on OpenRouter, the leading developer API marketplace, dropped from 40% a year ago to roughly nine to 13% today. With a reported 2026 IPO ahead, that trajectory is a serious problem.

What it means: Models built through illicit distillation strip out safety guardrails entirely. That is not theoretical. No international or domestic laws currently prevent this kind of model theft, and US chip export controls are essentially the main line of defense. This is a national security issue, not just a business story.

Microsoft is rolling out Copilot Tasks, an AI system that offloads work to Microsoft's cloud-based computer and browser while your local machine stays completely free.

Users describe tasks in plain language and Copilot handles the rest, from scheduling appointments and managing subscriptions to converting emails and attachments into finished slide decks. It can also plan events end-to-end, from booking venues to sending invitations, on a recurring schedule.

The system always requests permission before significant actions like payments or sending messages.

The catch: currently locked in a limited research preview for a very small group of testers. Expect a slow rollout.

What it means: Running tasks in the cloud rather than on local hardware is a smarter security architecture than what competitors are doing. Almost nobody can use this yet though. Get on the waitlist now and watch how quickly access opens up. The underlying capability is genuinely worth tracking.

Unlike Microsoft's tightly restricted preview, Anthropic introduced scheduled tasks for Claude Cowork and it's already live for all paid subscribers on Pro, Max, Team, and Enterprise plans.

Users can automate recurring workflows like daily briefings compiled from Slack and email, regular competitor analysis, and scheduled file organization. Everything runs inside a dedicated virtual machine, completely separate from the main operating system, with explicit permission required before any major actions.

The key difference from Microsoft: Cowork requires the computer and desktop app to stay actively running. If the machine sleeps, tasks get skipped and resume once it's back on.

What it means: Yuuuuup. Anthropic beat Microsoft to widespread availability by a wide margin. The computer-must-stay-awake limitation is real, but for anyone already on a paid Claude plan, this is actionable today. Automated daily briefings from Slack and email alone could meaningfully change how your mornings run.

Is this the biggest release ever for Perplexity? It could be.

Perplexity launched Perplexity Computer, a cloud-based agentic AI tool designed to function as a full digital coworker, operating entirely in the cloud with no local setup and no risk to personal files.

The system coordinates 19 different AI models simultaneously, assigning each subtask to the best-suited model: Claude Opus for logical reasoning, Gemini for deep research, GPT models for fact-based tasks, and other models for images and video.

It breaks any high-level goal into structured subtasks and delivers finished outputs like dashboards, apps, and complete presentations. Early comparisons to OpenClaw have been largely favorable.

Access is currently locked to the Perplexity Max subscription at $200 per month.

What it means: Perplexity needed a reason to exist as search got commoditized by OpenAI, Google, and Anthropic. This is their answer.

The multi-model architecture is genuinely smart and the cloud-based design solves real security concerns that local-run tools have not solved. Whether enterprise users add yet another $200 monthly subscription to their stack probably shapes Perplexity's entire future.