Do these 3 things before closing this tab:
1Scan for outdated or missing drivers - takes under a minute2Clear out junk files and repair common Windows errors3Fix the driver behind crashes, sound loss and screen glitchesAnthropic’s February 23, 2026 announcement alleges that DeepSeek, Moonshot AI and MiniMax used about 24,000 fraudulent accounts to generate more than 16 million exchanges with Claude, then used the results to improve competing AI systems. That is a serious accusation, but it is not proof that anyone stole Claude’s model weights, source code or complete architecture. The dispute is about alleged unauthorized model distillation: extracting useful capabilities from a hosted model through large-scale querying.
The headline’s irony is real only in a limited sense. Anthropic also uses AI feedback, synthetic data and principles drawn from public and third-party sources. Those practices deserve scrutiny, but the available evidence does not show that Anthropic committed the same alleged conduct—creating fraudulent accounts to harvest a competitor’s proprietary service.
As an Amazon Associate I earn from qualifying purchases.
What Anthropic says happened
In its February 23, 2026 announcement, Anthropic said it detected coordinated activity that was still active in some cases. The company alleges that the three named labs used roughly 24,000 fraudulent accounts and produced more than 16 million exchanges with Claude. Anthropic says the activity violated its terms of service and regional-access restrictions, and it called for coordinated action by AI companies, cloud providers and policymakers.
These figures, the attribution and the claimed intent are Anthropic’s account. The public announcement does not provide a complete account list, raw logs, a reproducible forensic method or an independent audit. No substantive response from DeepSeek, Moonshot or MiniMax appears in the sources available for this article.
#1 Best Overall
Distillation is a technique, not automatically theft
In ordinary model distillation, a larger “teacher” model produces answers, rankings, critiques, demonstrations or synthetic tasks. A smaller “student” model trains on those outputs to approximate useful behavior while reducing the cost of serving or developing it.
Companies commonly distill their own systems into faster, cheaper or specialized versions. Anthropic itself describes distillation as “widely used and legitimate” in some contexts. The disputed questions are who owns the teacher model, whether access was authorized, how accounts were obtained, how much data was collected and whether the activity targeted a rival’s commercially valuable capabilities.
What Anthropic alleges about each company
| Company | Exchanges Anthropic attributes to it | Capabilities Anthropic says were targeted |
|---|---|---|
| DeepSeek | More than 150,000 | Reasoning, rubric-based grading and censorship-safe responses for politically sensitive prompts |
| Moonshot AI | More than 3.4 million | Agentic reasoning, tool use, coding, data analysis, computer-use agents and computer vision |
| MiniMax | More than 13 million | Agentic coding, tools and orchestration |
The distribution matters. DeepSeek is the best-known name in the headline, but Anthropic’s own numbers identify MiniMax as the largest alleged source of traffic and Moonshot as substantially larger than DeepSeek.
Free tools Windows power users keep installed
One-click scans. No signup required.
DeepSeek
Anthropic says DeepSeek generated more than 150,000 exchanges and focused on reasoning abilities, grading rubrics that could support reinforcement learning, and ways to answer politically sensitive questions without triggering censorship. Anthropic also says some prompts asked Claude to explain the reasoning behind completed answers step by step. That is evidence of an attempt to create reasoning-like training data, not proof that the accounts accessed Claude’s hidden internal chain of thought.
Moonshot AI
Anthropic attributes more than 3.4 million exchanges to Moonshot. It says hundreds of fraudulent accounts used multiple access routes to target reasoning, coding, data analysis, computer use, vision and tool-enabled agents. The company further alleges that later activity tried to reconstruct Claude’s reasoning traces.
MiniMax
Anthropic attributes more than 13 million exchanges to MiniMax, mainly involving agentic coding, tool use and orchestration. It says nearly half of MiniMax’s traffic moved to a newly released Claude model within 24 hours of that model’s launch, and that the activity was detected before the MiniMax model Anthropic believes was being trained was released.
What evidence is public?
Anthropic says its investigation used IP-address correlations, request metadata, infrastructure indicators and timing compared with public product roadmaps. It also says some findings were corroborated by industry partners and that metadata linked certain accounts to specific researchers.
That supports a layered reading of the claim:
- Observed behavior: Anthropic reports unusual traffic patterns, synchronized requests and very large exchange volumes.
- Attribution: Anthropic concludes that particular labs controlled the accounts and infrastructure.
- Intent: Anthropic interprets the prompts as data collection for competing models.
- Legal status: Anthropic says its terms and regional restrictions were violated, but that statement is not a court finding.
The available material does not independently establish that a resulting model incorporated Claude-derived outputs, nor does it show that any accused company copied Claude’s weights, source code or proprietary training corpus.
Did DeepSeek copy Claude?
Not in the literal sense suggested by “copying the AI.” The allegation concerns repeated queries to Claude and the use of its outputs as training material. A student model can learn behaviors, skills and response patterns without containing Claude’s parameters or reproducing its full architecture.
That distinction also separates capability imitation from verbatim copying. The central dispute is about extracting know-how from a hosted service, not demonstrating a direct technical clone of Claude.
Why Anthropic calls it a national-security concern
Anthropic argues that safeguards built into a frontier model may not survive distillation. A student model could reproduce useful capabilities while omitting restrictions that limit assistance with cyber operations, disinformation, surveillance or other sensitive tasks. The company also argues that large-scale extraction could let restricted actors obtain capabilities indirectly, weakening the practical effect of export controls.
Those are Anthropic’s risk and policy arguments. They do not establish that DeepSeek, Moonshot or MiniMax used a distilled model for military or harmful purposes.
The “irony” of Anthropic’s own training methods
Constitutional AI and AI-generated feedback
Anthropic’s public descriptions of Claude’s Constitution and Constitutional AI explain a process in which a model critiques and revises answers against written principles. Anthropic has described training preference models from AI-generated evaluations rather than relying entirely on human harmlessness labels.
Synthetic data from Claude
In its account of Claude’s new constitution, Anthropic says Claude can help create synthetic conversations, constitution-aligned responses and rankings for future model training. That is model-generated training data, but it is described as being produced within Anthropic’s own development process.
Public and third-party principles
Anthropic says its constitution draws on the UN Declaration of Human Rights, principles from other AI laboratories, platform guidelines, non-Western perspectives and its own research. The constitution was released under CC0 1.0, which permits reuse.
These facts create a genuine tension about how AI companies learn from existing systems. They do not, by themselves, establish equivalence with allegedly fraudulent API access. The relevant comparisons are authorization, account authenticity, ownership of the teacher model, scale, contractual terms and competitive purpose. Using public principles or a company’s own model to generate synthetic data is materially different from allegedly operating thousands of accounts to harvest a rival’s hosted service.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Is the alleged conduct illegal?
Several legal questions must be kept separate:
- Terms of service: Anthropic says the activity breached its contractual rules and regional restrictions.
- Copyright: A terms-of-service breach does not automatically prove copyright infringement. Learning capabilities from outputs is not the same legal question as copying protected expression verbatim.
- Trade secrets: The public account does not establish that secret weights, code or confidential training data were obtained.
- Computer-access laws: Fraudulent accounts or evasion of access controls could raise separate issues, depending on the evidence and jurisdiction.
- Policy and export controls: Anthropic’s argument that distillation undermines restrictions is a policy position, not a determination that a national-security offense occurred.
Whether the evidence supports litigation or regulatory action remains an open question. The accused companies’ explanations, if issued, would be important to assessing attribution and intent.
Why the dispute matters beyond these three labs
- Frontier-model economics: If outputs can be harvested cheaply, the value of expensive training and safety work becomes harder to protect.
- Hosted-model security: Providers may need stronger identity checks, rate limits and behavioral monitoring without blocking legitimate users.
- Safety transfer: Distillation may preserve capabilities more reliably than safeguards, creating a policy concern even when the student model is smaller.
- Openness versus secrecy: Public principles and open research encourage progress, while closed APIs are easier to control but harder to scrutinize.
- Evidence standards: Public accusations should distinguish traffic observations from attribution, intent and legal conclusions.
What may happen next
AI providers are likely to experiment with stronger account verification, tighter regional controls, adaptive rate limits, detection of synchronized extraction patterns and cooperation with cloud platforms. Anthropic could also pursue contractual claims or seek government support. Those measures will face trade-offs: aggressive monitoring can affect privacy and legitimate research, while weak controls leave hosted models vulnerable to systematic harvesting.
Bottom line
Anthropic has described a technically plausible, industrial-scale distillation campaign and supplied specific figures for DeepSeek, Moonshot and MiniMax. The public record supports a serious allegation, not a final finding that any company stole Claude itself. The irony is that all major AI developers, including Anthropic, use forms of synthetic data, model feedback and knowledge drawn from existing work. The decisive difference is whether a competitor’s model was accessed lawfully and transparently—or allegedly extracted through fraudulent accounts and coordinated, unauthorized use.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Clear out junk files and repair common Windows errorsFree Scan →Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
