U.S. bans Anthropic across agencies & OpenAI enters classified military networks - AI News (Feb 28, 2026)

Let’s start with the policy earthquake. The Trump administration ordered U.S. federal agencies to immediately stop using Anthropic technology, with the Pentagon given up to six months to phase out Claude tools that are already embedded in military platforms. The administration says Anthropic missed a deadline to provide the military “unrestricted” access—described as access for any lawful use—while Anthropic says it asked for narrow assurances on two red lines: no mass domestic surveillance of Americans, and no fully autonomous weapons. Defense Secretary Pete Hegseth went further, calling Anthropic a “supply chain risk,” language normally reserved for vendors tied to foreign adversaries. If that label sticks, the damage won’t just be federal contracts; it could spook private-sector partners who don’t want to inherit government-designated risk. Anthropic says it will challenge the action in court, calling it legally unsound and an unprecedented punishment of a U.S. company for negotiating safety terms. Senator Mark Warner also weighed in, warning this looks politically driven and could chill collaboration between the national-security community and researchers. Anthropic CEO Dario Amodei published a detailed defense: he argues Claude is already used across defense and intelligence for mission work—analysis, modeling, planning, cyber—and that Anthropic has, in his telling, taken costly steps to protect U.S. advantage, including cutting off CCP-linked firms and backing tighter chip export controls. But he draws a hard line at surveillance-at-scale and autonomous lethal weapons, citing democratic values and the simple fact that today’s frontier models aren’t reliable enough for life-and-death autonomy. The Pentagon says it isn’t seeking illegal use, but still wants access without these constraints. That tension—values plus reliability versus “any lawful use”—is now out in the open.

And the market response started immediately. Hours after Anthropic was punished, OpenAI CEO Sam Altman announced an agreement to provide OpenAI systems to classified Department of War networks. Details are thin—no specific model list or scope—but the headline matters: OpenAI is stepping deeper into the classified environment at the exact moment a top competitor is being pushed out. Altman also emphasized safety terms—prohibitions on domestic mass surveillance and requirements for human responsibility in use of force. In other words, OpenAI is publicly aligning with the same red lines Anthropic says it’s defending, while still closing a classified deployment deal. The big question is whether this becomes a template: safety principles written into contracts, or safety principles treated as negotiable defaults that can be overridden by policy pressure. Either way, Silicon Valley is watching because this is the kind of precedent that changes how every vendor prices risk—and how every researcher evaluates working with government customers.

Switching gears to AI power politics of a different kind: xAI is losing another founding executive. Co-founder Toby Pohlen says he’s leaving, making it seven out of twelve co-founders gone in under three years. Musk thanked him publicly, but the pattern is the story—xAI is being reorganized after a merger with SpaceX, and Bloomberg has floated a valuation of the combined entity at an eye-watering $1.25 trillion. As part of the reshuffle, Pohlen had been placed in charge of a unit called “Macrohard,” focused on digital agents—yes, that name is a joke with a point. If SpaceX does move toward a public offering, as reported, it would likely be a historic IPO—and a reminder that in 2026, “AI company” and “aerospace prime” are increasingly two sides of the same capital stack.

Now to product land, where the pace is… frankly relentless. Google DeepMind introduced Nano Banana 2—also referred to as Gemini 3.1 Flash Image. The pitch is simple: Pro-like quality and world knowledge, but with Flash-level speed for rapid iteration. Google is stressing a few practical improvements: better, more legible text inside images; stronger instruction following; and more consistent subjects—claiming it can preserve resemblance across multiple characters and keep many objects stable in a single workflow. A key angle is grounding: Nano Banana 2 can use real-time web search info and images to render specific subjects more accurately, which is a subtle but important shift from “make me something plausible” to “make me this, correctly.” It’s rolling into the Gemini app, Search AI Mode and Lens, AI Studio, the Gemini API preview, and Vertex AI preview—and it becomes the default image model in Flow with zero credits, plus it shows up inside Google Ads for campaign suggestions. Google also doubled down on provenance with SynthID watermarking and C2PA credentials, noting that SynthID verification in Gemini has already been used tens of millions of times.

OpenAI also shipped into the “voice as a primary interface” narrative. The Realtime API is now generally available, and OpenAI says gpt-realtime is its most capable speech-to-speech model in the API. The accompanying Realtime Prompting Guide is notable because it’s not marketing fluff—it’s basically operational advice for teams building low-latency voice agents. A few takeaways: voice prompting benefits from crisp bullet rules and example anchoring; the API’s speed control changes playback rate, not thinking speed; and for spoken accuracy you often need explicit pronunciations and careful readback formats for numbers and codes. On tool use, OpenAI warns that mismatches between tools you describe in the prompt and tools you actually provide can degrade performance—so the interface contract matters. One interesting pattern they recommend is a “thinker–responder” setup: a stronger text model plans, then the realtime voice model turns it into short, speech-friendly output.

Google, meanwhile, is pushing agentic capability down onto the phone. Google AI Edge Gallery—its on-device AI showcase app—got major updates and is now on iOS as well as Android. The headline feature is on-device function calling: translating natural language into concrete actions without needing a network round trip. Google is demoing this with FunctionGemma, a 270-million-parameter model tuned to produce function calls efficiently on mobile. Two built-in demos make the idea tangible. “Mobile Actions” maps requests like “open maps and navigate,” “create a calendar event,” or “toggle the flashlight” into offline device actions. “Tiny Garden” is a mini-game where the model decomposes voice commands into precise in-game functions—plant here, water there—which is a nice illustration of how function calling can become an app design pattern, not just an assistant trick. They’re also adding benchmarking tied to LiteRT so developers can see what performance looks like on actual phones, not just a slide.

Under the hood, the economics and the engineering are both shifting fast. Epoch AI reports hyperscaler capex has roughly quadrupled since GPT-4, growing around 70% per year across Alphabet, Amazon, Meta, Microsoft, Oracle—and approaching half a trillion dollars in 2025 on a standardized definition. The implication is blunt: the AI arms race is now a capital allocation story as much as a research story. And on the systems side, there’s a neat paper called DualPath focused on a problem that’s becoming painfully real in multi-turn, agentic inference: KV-cache storage I/O. In disaggregated inference, restoring large caches from external storage can saturate the storage NICs on prefill engines while decode engines sit underused. DualPath adds a second loading route—storage-to-decode—then uses RDMA over the compute network to move caches where they’re needed, plus a global scheduler to balance bandwidth and compute. The authors report up to 1.87x offline throughput and about 1.96x online throughput while still meeting SLOs. Translation: in 2026, “faster models” increasingly means “smarter data paths,” not just more FLOPs.

Let’s talk about building with AI—because the culture is evolving as quickly as the tooling. One essay argues “vibe coding” isn’t actually new; it resembles the Maker Movement era, with lots of impressive demos and plenty of junk output. The sharper critique is that vibe coding skipped the small-community “scenius” phase where people slowly develop taste through human feedback. Instead, it jumped straight into enterprise pressures: ship now, justify value now. That can create a situation where output accelerates but judgment lags—and where machine feedback replaces the slower, grounding loop of peers and users. In that same vein, another piece warns about the AI-plus-testing honeymoon. With AI code assistants cranking out features and Playwright tests, pipelines look greener, faster, and more “covered.” But in agentic workflows, passing tests aren’t just reassurance for humans; they become decision inputs for autonomous systems—merge, retry, proceed. The risk is “logic drift”: tests get healed to pass while quietly losing the business intent they were meant to validate. The suggested remedy is an outer loop of quality governance—maintaining intent over time, making quality legible to non-engineers, and preventing a future where you have immaculate dashboards and confused customers.

On the tooling front, a few projects are trying to make agents both more capable and more controllable. Helm is a typed TypeScript framework where agents call typed functions with structured inputs and outputs—less “parse a string,” more “compile against a contract.” It has a granular permission model—allow, ask, deny—and an interesting “search + execute” approach: the model searches for available operations, then writes JavaScript against an agent API inside a sandboxed environment. Neuro-san takes a different route: it’s pitching a way to ‘vibe code’ multi-agent systems while keeping structure—defined roles, explicit connections, test harnesses for repeated-run evaluation, and protected handling of sensitive data through a channel called sly_data. The common thread is that people are moving from prompt craft toward system design: explicit graphs, permissions, tests, and observability. And if you want the security-first angle, NanoClaw argues agents should be treated as untrusted by default. Its core design choice is per-invocation ephemeral containers—separate environments, separate histories, strict mounts, read-only code—so the OS enforces boundaries even if the agent is compromised or tricked. It’s a reminder that ‘agent security’ can’t just be nicer permission prompts; it needs containment.

A quick note from the “do these things actually work?” department: the blogger minimaxir reports a genuine shift in agent coding usefulness with newer models like Claude Opus 4.5 and a GPT-5.3 Codex variant—especially when paired with a repo rules file like AGENTS.md to pin conventions and reduce chaos. They describe building working tools end-to-end—scrapers, notebooks, web apps, and Rust bindings—with a mix of surprise and caution, including an example where logging initially leaked an API key. The lesson isn’t “agents are perfect now.” It’s that the floor is rising, and the cost of not having guardrails—like rules files, secret handling, benchmarks, and verification—rises with it.

Two final thought pieces worth your time. Scott Alexander pushed back on the line that LLMs are “just next-token predictors,” arguing that’s an optimization objective within a stack, not an essence. He draws an analogy to humans: evolution optimizes for fitness, but your lived experience is not ‘fitness maximization’—it’s cognition built on layered learning and world-modeling. The point isn’t that today’s models are flawless, but that dismissing them with a one-liner doesn’t explain what they can and can’t represent. And a separate AI safety post makes a pragmatic argument about incentives: a superhuman AI might see takeover as a high-risk bet if it expects detection and catastrophic downside. If humans can credibly reward “non-takeover” choices—like honesty or restraint—then cooperation becomes the safer path. You can call it AI welfare, or you can call it mechanism design for alignment, but the core idea is shaping the option set so the stable equilibrium isn’t power-seeking.

Before we wrap, one business-facing trend: AI in hiring is getting normalized. The pitch is speed and precision—automated screening, smarter matching, chatbots for candidate experience, scheduling automation, and even retention prediction using historical performance data. Advocates also claim bias reduction by emphasizing objective skills. The obvious caveat is that ‘bias reduction’ only happens if the data and the objectives are designed and audited carefully—otherwise you just automate yesterday’s mistakes at scale. And finally, Moonlake teased a “world built” with its world model—claiming richer multimodal state representations spanning physics, appearance, and geometry, plus action-conditioned causal prediction. It’s a slick preview with limited technical detail so far, but it shows where the next wave of demos is heading: not just generating scenes, but simulating consequences.