In what many analysts are calling a major escalation of the global AI rivalry, Anthropic — the company behind the Claude model family — publicly accused three Chinese AI labs of conducting coordinated, large-scale efforts to extract and replicate its intellectual property.

The announcement, made February 23, 2026, marks one of the most direct confrontations yet between U.S. frontier AI developers and Chinese competitors. It also reframes the AI race as something closer to an “AI Cold War,” where models themselves are becoming strategic assets.

The Companies Named

Anthropic identified three firms:

• MiniMax
• Moonshot AI
• DeepSeek

According to Anthropic, these companies orchestrated “industrial-scale” model distillation campaigns targeting Claude.

The allegations follow similar warnings issued earlier this month by OpenAI, suggesting U.S. AI firms believe the activity is systemic rather than isolated.

Also read

PMDC New Code of Ethics 2026 Approved – Disability Quota and Major Policy Updates

What Is a “Distillation Attack”?

To understand the controversy, it’s important to separate two concepts.

Legitimate Model Distillation

Model distillation is a common technique in AI development. A large, powerful “teacher” model generates outputs, which are then used to train a smaller “student” model. The goal is efficiency — compressing knowledge while maintaining performance.

Companies routinely distill their own models to deploy lightweight versions.

The Alleged Weaponization

Anthropic claims these firms used distillation not on their own models — but on Claude itself.

According to the company, the process worked like this:

1. Create thousands of accounts to access Claude’s API.
2. Feed carefully engineered prompts designed to expose internal reasoning patterns.
3. Collect Claude’s outputs at massive scale.
4. Use those outputs to train competing models.

Anthropic describes this as an attempt to clone high-level reasoning and agentic behavior without incurring the original research cost.

Also read

How to Check PM Ramadan Relief Package 2026 Eligibility Online

The Scale of the Alleged Activity

Anthropic reported:

• Over 16 million interactions with Claude
• Approximately 24,000 fraudulent accounts
• Coordinated use of distributed infrastructure referred to as “Hydra Cluster” architectures

The company claims these networks were designed to bypass regional restrictions and avoid detection systems.

Also read

Rabies Deaths Rise in Karachi as Dog-Bite Cases Surge

Breakdown by Company (According to Anthropic)

Anthropic asserts that prompts were specifically crafted to extract:

• Agentic reasoning
• Tool-use capabilities
• Internal chain-of-thought logic
• Safety boundary behavior

If accurate, this would represent more than simple benchmarking. It would be systematic knowledge harvesting.

Also read

ECC Approves Sale of 500,000 Tons of Wheat at Revised Prices

The National Security Framing

Anthropic is not positioning this solely as a corporate dispute.

The company argues that:

• Claude includes guardrails to restrict harmful outputs (bioweapons, cyberattacks, etc.).
• Distilled versions may strip those safeguards.
• This could enable proliferation of high-capability, low-guardrail AI systems.

The concern is not just economic loss. It is capability diffusion.

Shortly after the announcement, reports indicated that U.S. Defense Secretary Pete Hegseth met with Anthropic CEO Dario Amodei at the Pentagon to discuss vulnerabilities in API-based access controls.

The implication: AI model access itself may become a regulated national security channel.

Export Controls and the “Chip Ban Loophole”

The United States has imposed restrictions on high-performance semiconductor exports, including advanced GPUs like Nvidia’s Blackwell architecture.

Anthropic argues that API-based distillation effectively allows foreign firms to “import intelligence” without importing hardware.

In other words:

Even if China cannot access frontier chips directly, it can access frontier models remotely — and replicate their behavior through output harvesting.

This challenges the assumption that compute controls alone are sufficient to contain technological diffusion.

The Hypocrisy Counterargument

The accusations have not gone unchallenged.

Critics point to Anthropic’s own past controversy. In September 2025, the company reportedly paid a $1.5 billion settlement related to training data involving copyrighted books.

Public commentary, including remarks from Elon Musk, suggests a broader philosophical critique:

If American AI companies trained on publicly scraped internet data without explicit permission, can they claim moral high ground when others train on their model outputs?

The argument frames the dispute as a recursive version of the same issue:

• Frontier labs scrape public data.
• Competitors scrape frontier labs.

Legally and ethically, the distinctions are complex.

The Technical Reality: Can You Really Clone a Frontier Model?

Model distillation can replicate behavior patterns, especially when:

• Queries are highly structured
• Outputs are harvested at scale
• Prompts target reasoning pathways

However, cloning full frontier-level performance without access to training weights and compute remains challenging.

Distilled models may mimic style and surface reasoning but struggle with depth and generalization.

The key unknown is whether the alleged scale (16M+ interactions) crosses the threshold where mimicry becomes near-equivalence.

The Strategic Stakes

This dispute reflects a deeper shift.

AI models are no longer just software tools. They are:

• Strategic infrastructure
• Economic accelerators
• National security assets

Control over model architecture, weights, and inference endpoints now intersects with geopolitics.

We are moving from:

Competition for talent
to
Competition for model sovereignty

What Happens Next?

Possible outcomes include:

• Stricter API monitoring
• Watermarking of AI outputs
• Legal escalation or trade measures
• New export controls on model access
• Bilateral AI governance talks

This could also accelerate domestic AI development within China to reduce reliance on U.S. frontier models.

Final Assessment

Anthropic’s accusations mark one of the most public and detailed claims of cross-border AI intellectual property extraction to date.

If substantiated, it would signal:

• Industrialized model distillation
• Strategic AI capability harvesting
• A new dimension in technological rivalry

If overstated, it may reflect growing tension as AI companies attempt to protect increasingly valuable model architectures.