Anthropic says new Claude AI safeguards have reduced unauthorised model data extraction by up to 99%, strengthening protection against AI copying attacks.
SAN FRANCISCO: Anthropic has introduced stronger Claude AI safeguards to prevent rival developers from copying its artificial intelligence models, reporting a 90% to 99% reduction in data extracted through unauthorised attacks.
Announced on October 12, 2026, the measures target attempts to reproduce Claude’s advanced capabilities without permission, addressing growing concerns about intellectual property, cybersecurity and AI competition.
What are AI model copying attacks?
Known as distillation attacks, these involve extracting responses from a powerful AI model and using them to train another system with similar capabilities.
Although model distillation is a legitimate technique when authorised, Anthropic says some developers have used fraudulent accounts and automated requests to extract Claude’s protected reasoning.
These attacks can allow competitors to reproduce advanced AI capabilities without investing equivalent resources in research and development.
How do Anthropic’s new safeguards work?
The company has introduced several protective measures to make unauthorised extraction more difficult.
The updated Claude AI safeguards include tighter restrictions on access to internal reasoning, automated detection of suspicious requests and stronger controls over account activity.
Anthropic has also introduced a two-stage detection system designed to identify attempts to reveal protected reasoning.
According to the company, the detection system has a false-positive rate below 0.1%, helping minimise disruption to legitimate users.
Additional measures restrict repeated extraction attempts and allow the company to suspend accounts linked to suspicious activity.
How effective are the new protections?
Anthropic claims its defences have reduced the amount of reasoning data extracted by unauthorised developers by approximately 90% to 99%.
The company says the volume of protected reasoning collected through these attacks has fallen by a factor of 10 to 100.
Tens of thousands of accounts associated with suspected distillation activities have also been banned.
However, Anthropic acknowledges that the improvements may not be permanent, as attackers could develop new methods to bypass its protections.
Why does this matter for the AI industry?
The announcement comes as competition intensifies among major artificial intelligence developers.
Anthropic argues that unauthorised model copying could enable the creation of powerful AI systems without equivalent safety protections.
Stronger defences could help technology companies protect proprietary research while reducing opportunities for misuse.
However, stricter security controls may also affect legitimate developers, particularly those integrating Claude into applications.
Impact to expect
The stronger Claude AI safeguards could encourage other AI companies to improve model security and monitor suspicious activity. Businesses using AI services may benefit from greater protection, although developers could face tighter access controls and additional technical requirements.
For the wider technology industry, the announcement highlights the growing importance of balancing innovation, competition and responsible AI development.
