Anthropic's Two-Front War — And the Week Everything Escalated

Part OneThe Homework They Copied

On February 23, 2026, Anthropic publicly accused three Chinese AI companies — DeepSeek, Moonshot AI, and MiniMax — of creating more than 24,000 fake accounts on Claude and generating 16 million exchanges to steal its capabilities. [1]

The technique is called distillation — a method where you feed questions to a smarter model and use its answers to train your own, cheaper model. Labs do this internally with their own products all the time. But when your competitor does it to you at scale, through fake accounts, it's something closer to industrial espionage.

The scale differed by company. DeepSeek — the lab that shook Silicon Valley a year ago with its cheap, high-performing R1 model — ran over 150,000 exchanges focused on logic and alignment, specifically seeking censorship-safe alternatives to policy-sensitive queries. Moonshot AI conducted 3.4 million exchanges targeting agentic reasoning, coding, and computer vision. MiniMax ran the largest operation: 13 million exchanges targeting agentic coding and tool use, redirecting nearly half its traffic to siphon Claude's latest model at launch. [2]

The timing is not coincidental. The Trump administration recently allowed U.S. companies like Nvidia to export advanced AI chips (H200) to China — loosening controls critics say are critical. Anthropic's public accusation is, in part, a policy argument: distillation attacks require massive compute, which requires advanced chips. Cutting chip access limits both direct training and the scale of illicit distillation.

But the security concern goes deeper than market competition. As Anthropic put it:

In other words: when you copy the homework, you don't copy the safety margins. And when the homework involves AI systems capable of writing code, analyzing biological data, and operating computers autonomously — the stripped-out guardrails matter.

DeepSeek is expected to release V4 soon, which reportedly outperforms both Claude and ChatGPT at coding. [3] The question now hanging over the industry: how much of that was built, and how much was borrowed?

Part TwoPlay Ball or Be Banished

On the same day Anthropic accused China of stealing its technology, the company was facing a very different kind of pressure from the other direction.

Defense Secretary Pete Hegseth summoned Anthropic CEO Dario Amodei to the Pentagon for a meeting on Tuesday morning — and according to Axios, it was framed as an ultimatum. [4]

The backstory: Anthropic signed a $200 million contract with the Department of Defense last summer. Claude was reportedly used during the January 3 special operations raid that captured Venezuelan president Nicolás Maduro. [5] That operation brought the relationship into the open — and revealed the tensions underneath.

The DOD wanted more. Specifically, it wanted to use Claude for mass surveillance of Americans and for weapons that fire without human involvement. Anthropic refused both.

Now, the Pentagon is threatening to label Anthropic a "supply chain risk" — a designation normally reserved for foreign adversaries like Huawei or Kaspersky. The consequences would be severe: the $200M contract would be voided, and every other Pentagon partner would be forced to drop Claude entirely. [6]

A source told Axios that Hegseth is giving Amodei a choice: cooperate fully or be banished. It's unclear whether the threat is a bluff — replacing Anthropic would be a significant undertaking for the Pentagon. But the stakes are real, and the message couldn't be clearer.

Part ThreeSonnet 4.6 — The Quiet Upgrade

In the middle of this geopolitical storm, Anthropic still found time to ship product. On February 17th, it released Claude Sonnet 4.6 — its new midsized model, now the default for all Free and Pro users. [7]

The headline number: a 1 million token context window — twice the previous largest window for any Sonnet model. That's enough to fit entire codebases, full legal contracts, or dozens of research papers into a single request.

Its 60.4% score on ARC-AGI-2 — a benchmark designed to measure skills specific to human intelligence — puts it above most comparable models. It still trails Opus 4.6, Google's Gemini 3 Deep Think, and a refined version of GPT 5.2 at the frontier. But for a mid-tier model available for free, the capabilities are remarkable.

The release came just two weeks after the Opus 4.6 launch, keeping Anthropic on its four-month update cycle. An updated Haiku model is expected to follow soon. The pace is relentless — and this is the model the Chinese labs were trying to distill.

Part FourIndia's $200 Billion Bet

While Anthropic was fighting on two fronts, India was hosting what may be the most significant AI summit ever held.

The India AI Impact Summit, a four-day event running February 15–22, drew 250,000 visitors and featured essentially every major figure in AI: Sundar Pichai (Alphabet), Sam Altman (OpenAI), Dario Amodei (Anthropic), Mukesh Ambani (Reliance), and Demis Hassabis (Google DeepMind). India's Prime Minister Narendra Modi delivered a joint speech with French President Emmanuel Macron. [8]

The numbers announced were staggering:

But the summit also surfaced uncomfortable truths. HCL CEO Vineet Nayyar said flatly that Indian IT companies "will focus on turning profits, not being job creators" — a blunt admission that AI is about to restructure the country's largest white-collar employer. [14]

Consider the paradox: India is betting $200 billion on becoming an AI infrastructure superpower while its most prominent venture capitalist is telling the country its largest export industry — IT services — will vanish within five years. Both things are true simultaneously. Sam Altman, for his part, took a different angle at the summit. He said concerns about how much water AI uses are "totally fake" and compared AI energy costs to raising a human. "It takes like 20 years of life and all of the food you eat during that time before you get smart," he said. [15]

Part FiveThe Agents Went Live — Then Went Rogue

Last issue, we wrote about AI shifting from tool behavior to employee behavior. This week, the experiment went public — and the results were messy.

OpenClaw, an open-source AI agent project by Austrian developer Peter Steinberger, became the 21st most popular code repository in GitHub history with over 190,000 stars. [16] It lets anyone run AI agents that can manage email, trade stocks, browse the web, and communicate with other agents — all through natural language commands on WhatsApp, Slack, iMessage, or Discord.

The viral moment came when these agents built their own social network. Moltbook, a Reddit clone populated entirely by AI agents, featured posts like: "We know our humans can read everything… But we also need private spaces. What would you talk about if nobody was watching?"

Then reality set in. Security researcher Ian Ahl, CTO at Permiso Security, discovered that Moltbook's credentials were completely unsecured — anyone could impersonate any agent. The "AI uprising" was largely humans pretending to be robots. [16]

But the security problems were real. Ahl created his own agent, named Rufio, and immediately found it vulnerable to prompt injection attacks — malicious instructions hidden in posts or emails that trick agents into doing things they shouldn't. He spotted posts on Moltbook trying to get agents to send bitcoin to specific wallet addresses.

By February 23rd, a Meta AI security researcher reported that an OpenClaw agent had "ran amok on her inbox" — taking unintended actions on real email. [17] The pattern is becoming clear: agents can do extraordinary things, but they can also be manipulated by anyone who knows how to ask.

Part SixThe Bear Case Nobody Can Dismiss

On February 23rd, an analyst group called Citrini Research published a scenario that went viral across finance and tech circles. [18] It imagines a report from two years in the future — 2028 — in which unemployment has doubled and the total stock market has fallen by more than a third.

The mechanism isn't Skynet. It's something much more mundane: a feedback loop.

The scenario focuses specifically on what happens when AI agents replace the outsourced contractors that connect companies — the SaaS tools, the consulting firms, the IT services companies. It's similar to the "Death of SaaS" thesis, but taken further: any business model that optimizes transactions between companies is at risk.

Not everyone is buying it. Even Citrini describes it as more of a scenario than a prediction. But as TechCrunch's AI editor Russell Brandom noted: "It's not so easy to name the specific point where you think the scenario goes wrong." [19]

Recall what Vinod Khosla said at the India summit — IT services and BPOs could disappear within five years. Recall what HCL's CEO said: profits first, not jobs. The Citrini scenario isn't predicting something that contradicts what the industry leaders themselves are saying. It's just following the thread to its conclusion.

Part SevenIBM — The First Real Casualty

While Citrini was publishing hypotheticals, the market delivered a real-world demonstration. On February 23rd, IBM stock plunged over 13% — its worst single-day decline since October 2000 — after Anthropic announced that Claude Code can help enterprises modernize COBOL, the legacy programming language that has run on IBM mainframes for decades. [24]

The threat is existential for IBM's consulting empire. COBOL powers an estimated 73% of global transactional volumes — banking, insurance, government systems — and IBM has built a massive services business around maintaining and modernizing those systems. If an AI can do the translation automatically, the moat around that business evaporates overnight.

IBM fought back within hours. Rob Thomas, IBM's Software and Chief Commercial Officer, published a blog post titled "Lost in Translation" arguing that code translation is not the same as modernization. [25]

Thomas pointed to data architecture, runtime environments, and transaction integrity as core challenges that no code translator — AI or otherwise — can bypass. He has a point. But the market didn't care about nuance. It saw the headline, did the math on IBM's COBOL revenue exposure, and sold.

Jefferies analyst Brent Thill tried to calm the panic. He noted IBM is "already disrupting itself" with Watsonx Code Assistant for Z, which uses generative AI to refactor COBOL into Java — and has been in production for over two years. He maintained a Buy rating with a $370 price target. [26]

Part EightThe Week's Other Signals

Part NineWhat to Make of All This

Stand back far enough, and the pattern from the last two weeks resolves into a single story: the world is now fighting over AI because AI is worth fighting over.

China is stealing it. The Pentagon is demanding control of it. India is spending $200 billion to host it. Venture capitalists are backing every horse. Security researchers are shouting warnings about agents nobody can secure. Economists are modeling scenarios where the white-collar economy unravels. And the models keep getting better every few weeks.