Anthropic flags industrial scale distillation attacks

Industrial scale distillation attacks have emerged as a new flashpoint in the AI race, following allegations that a leading AI model was systematically mined to accelerate rival model development.

In a public disclosure, the company behind the Claude model stated that three AI laboratories orchestrated large-scale efforts to extract its capabilities. According to the company, the activity involved more than 16 million exchanges conducted through over 24,000 fraudulent accounts.

The alleged campaigns were described as coordinated attempts to replicate advanced capabilities rather than legitimate usage.

What industrial scale distillation attacks mean

Distillation is a recognised AI technique. It allows a smaller model to learn from the outputs of a more advanced system. Many AI companies use distillation internally to create lighter or more cost-efficient versions of their own models.

The issue here, according to the disclosure, is scale and intent.

The company claims that the process was used externally to copy strengths such as:

• Agentic reasoning

• Coding capability

• Tool usage

• Structured problem solving

By repeatedly prompting the model in specific formats, the actors allegedly gathered high-quality outputs that could accelerate their own training cycles and reduce development costs.

In effect, industrial scale distillation attacks shift the competitive equation. Instead of building foundational capabilities independently, rivals may narrow the gap by harvesting model outputs at scale.

Three separate campaigns, varying intensity

The disclosure identified three laboratories: DeepSeek, Moonshot AI and MiniMax.

Each campaign differed in scope and execution.

• One operation reportedly generated more than 150,000 exchanges focused on reasoning patterns. Some prompts allegedly requested detailed, step-by-step explanations to capture structured logic.

• Another campaign scaled significantly higher, with more than 3.4 million exchanges targeting coding, computer use agents and data analysis.

• The largest campaign exceeded 13 million exchanges. It allegedly shifted activity to newly released model versions within 24 hours of updates, suggesting active monitoring and rapid adaptation.

The company stated that attribution relied on IP address correlation, metadata analysis, infrastructure indicators and confirmations from industry partners.

Fraudulent accounts and proxy networks

A central element of the alleged industrial scale distillation attacks was account manipulation.

The disclosure claims that more than 24,000 fraudulent accounts were created to generate traffic that blended with legitimate usage. In one case, a single proxy network allegedly managed over 20,000 accounts simultaneously.

The model provider does not offer commercial access in China. According to the company, proxy services were used to bypass regional restrictions and sustain extraction activity.

This introduces a broader concern: when access controls become part of geopolitical strategy, circumvention efforts may escalate in parallel.

Security and export control implications

Beyond commercial rivalry, the disclosure raised national security considerations.

Frontier AI models contain safeguards designed to prevent misuse in areas such as cyber attacks, disinformation and biological threats. Illicitly distilled models, the company argued, may not inherit those protections.

The issue also intersects with export controls. The company has supported measures aimed at preserving technological advantage. Industrial scale distillation attacks, it claims, weaken those controls by allowing foreign laboratories to narrow capability gaps without independent innovation.

In other words, the debate is no longer only about model performance. It is about governance, compliance and strategic leverage.

Response measures and industry coordination

In response, the company stated it has strengthened its defences through:

• Advanced detection systems

• Behavioural fingerprinting tools

• Classifiers to identify coordinated activity

It is also sharing intelligence with other AI labs, cloud providers and authorities in an effort to build a coordinated response framework.

The message is clear: protecting advanced models now requires cross-industry collaboration.

A new front in the AI competition

Industrial scale distillation attacks highlight a structural challenge in the AI ecosystem. As models become more capable and economically valuable, the incentive to replicate them grows.

The competition is shifting.

It is no longer only about training larger models or releasing smarter systems. It is also about protecting intellectual capital, enforcing usage policies and maintaining control over strategic technologies.

For policymakers and industry leaders alike, this episode underscores a difficult reality: the AI race is not just about innovation. It is about defence.

Read More: 

Solitaire introduces C-Series interactive panels for classrooms and workplaces

CrowdStrike 2026 Global Threat Report flags AI surge

SigTuple AS76 digital morphology analyser launched

Satcom Aurva distribution agreement targets DPDPA compliance