Pentagon presses Anthropic guardrails & Distillation attacks and AI theft - Tech News (Feb 26, 2026)

Let’s start with the U.S. Defense Department and Anthropic. Reporting says the Pentagon has given Anthropic a Friday deadline to agree to broader military use of its models, and it’s floating serious pressure tools—everything from invoking the Defense Production Act to potentially treating Anthropic as a supply-chain risk. The dispute appears to revolve around “usage policy constraints”: Anthropic doesn’t want its models used for certain autonomous weapons applications or mass domestic surveillance, and the Pentagon wants fewer carve-outs. Anthropic says it’s in good-faith talks and argues it can support national security within what models can do reliably and responsibly.

In parallel, another security fight is heating up: distillation and model extraction. Anthropic claims actors tied to Chinese AI firms generated massive volumes of Claude conversations via thousands of fake accounts—essentially siphoning outputs to train cheaper competing models. OpenAI and Google have been issuing similar warnings. Distillation isn’t inherently shady—labs legitimately compress their own models all the time—but the concern here is adversarial copying at industrial scale, potentially recreating capabilities without the original safeguards. The broader message from U.S. labs is clear: the AI “cold war” is increasingly about data pipelines and access controls, not just bigger GPUs.

Speaking of GPUs, Meta’s spending streak continues. A report says Meta has struck a deal to buy AMD’s newest MI450 AI chips in a multi-gigawatt agreement that could exceed one hundred billion dollars, with shipments ramping from an initial tranche later this year. The eye-catching piece is the structure: AMD issued Meta a performance-based warrant that could translate into a meaningful equity stake if purchase milestones are hit. It’s another sign that hyperscalers want optionality—multiple chip suppliers, multiple contract levers—while they scale data centers fast enough that power, racks, and delivery schedules become strategic constraints.

On the other side of the Pacific, Reuters says DeepSeek is changing its pre-release playbook for its next flagship model update, V4. Instead of giving Nvidia and AMD early access for tuning—pretty common in the industry—DeepSeek reportedly prioritized domestic Chinese suppliers, including Huawei, effectively giving them a head start on optimization. The practical effect may be modest in the short term, but symbolically it’s big: model releases are now also hardware-alignment events, and “who gets early access” is starting to look like industrial policy by other means.

Now to crypto, where one lawsuit is spawning a much wider set of claims online. A federal case in Manhattan tied to the TerraUSD collapse alleges that Jane Street received material nonpublic information about Terraform Labs’ liquidity actions—reportedly via a private chat group linked to a former Terraform intern who later joined Jane Street. The complaint points to a key moment on May 7, 2022: Terraform pulled a large amount of UST liquidity from Curve, and—according to the filing—an allegedly Jane Street–linked wallet withdrew tens of millions shortly after, before public disclosure, adding selling pressure ahead of the depeg. Jane Street has called the suit baseless and argues losses stem from Terraform’s own misconduct. Separate litigation also targets Jump Trading, suggesting the fallout may not be limited to one firm.

That lawsuit is being used to prop up a second, more speculative narrative: that Bitcoin’s price action has been “suppressed” through repeated, sharp sell-offs around the U.S. market open, often knocking out leveraged longs and then snapping back. Some analysts on social media have flagged recurring intraday patterns, and the thread making the rounds argues those moves paused when the Terraform-related lawsuit became public, then resumed. It also points to Jane Street’s disclosed holdings in BlackRock’s spot Bitcoin ETF and reminds listeners that 13F filings show long positions but not the derivatives that could hedge—or even flip—net exposure. Important caveat: patterns aren’t proof, and the lawsuit is still an allegation. But the episode highlights a real transparency gap in modern markets: the public often sees fragments of positioning, not the full risk book.

Switching to social platforms: X has now open-sourced key components of its “For You” feed system under an Apache-2.0 license. The code shows a real-time recommender pipeline that blends posts from accounts you follow with out-of-network discovery, then ranks them with a Grok-based transformer model. It’s a modern architecture: candidate retrieval via embeddings, enrichment with metadata, multiple scoring passes, and post-ranking filters to remove duplicates, spam, and blocked content. Beyond the tech, it’s an unusual move in a world where ranking systems are typically guarded—so expect researchers to scrutinize the mechanics, and competitors to borrow ideas where they can.

In developer-land, Cloudflare is leaning hard into the idea that AI plus tests can dramatically accelerate rewrites. The company says it rebuilt about 94% of the Next.js 16 API surface as an experimental Vite-based drop-in called vinext, aimed at smoother deployment on Cloudflare Workers—and it claims much of the code was AI-assisted, with the cost of tokens coming in around a thousand dollars. The more strategic takeaway is the “tests as a moat” theme: if you can validate compatibility with thousands of automated tests, you can replicate behavior quickly without inheriting years of internal baggage. It’s a shot across the bow for any framework vendor relying on complexity and version churn as an implicit lock-in.

Two quick AI-and-creators updates. First, Sony AI published research toward “musical integrity”: training-data attribution techniques, better short-clip version matching, and a reality check on watermarking. One notable finding: watermark schemes can fail completely when audio passes through neural audio codecs—suggesting watermarking won’t work in isolation; it may need cooperation from the codec layer. Second, Adobe is rolling out a beta feature called Quick Cut in Firefly video editor, designed to auto-assemble a rough first draft from b-roll and prompts. It’s not promising a finished edit—more like a fast starting point—an approach that’s quietly becoming the standard pattern for creative AI tools.

Finally, a pair of “physical world” stories. Wayve, the London autonomous-driving startup, raised $1.2 billion at an $8.6 billion valuation, pitching a licensing model to automakers rather than running its own robotaxi fleet. It says it expects a driverless-taxi commercial trial with Uber in London this year, and supervised consumer vehicles by 2027. And Alphabet’s Intrinsic—focused on making industrial robots easier to program—is moving closer into Google while remaining a distinct unit, pairing up more directly with DeepMind, Gemini models, and Google’s cloud stack. The common thread: AI is migrating from screens to streets and factory floors, and the business models are still being stress-tested in real time.

As a bonus note for science: MIT researchers published new AI tools that can track and label neurons in small, transparent animals even while their bodies move and deform. They report dramatic speedups and high accuracy, including methods that can discover cell types without supervision. If that holds up across labs, it’s one of those quiet breakthroughs that changes what’s feasible—not by a single flashy demo, but by removing a bottleneck researchers have been fighting for years.