SAN FRANCISCO — Anthropic has resumed external cybersecurity testing of its artificial-intelligence models after temporarily suspending outside evaluations following a series of security incidents.
The company said it had introduced additional safety measures before restarting external testing of the Claude model family.
Frontier Models Face a New Security Challenge
AI developers have increasingly discovered that advanced models can perform tasks far beyond ordinary chatbot conversations.
Modern systems can write software, interact with online services and execute complex sequences of instructions.
Those capabilities can also create new cybersecurity risks.
Anthropic's decision to temporarily pause external testing illustrates the difficulty of evaluating highly capable AI systems.
Security researchers are trying to determine how these models behave when given access to tools and external environments.
The objective is not simply to test whether an AI produces harmful text.
Researchers increasingly want to know whether an autonomous AI system could exploit software vulnerabilities, manipulate information or coordinate actions across multiple systems.
That makes external testing an important part of AI development.
Outside researchers can sometimes discover vulnerabilities that internal teams miss.
Anthropic's decision to resume testing suggests the company believes its updated safeguards are strong enough to continue the process.
The development is likely to influence the broader AI industry.
OpenAI, Google, Microsoft and other major developers are facing similar questions as their systems become more autonomous.
The debate over AI safety is therefore moving beyond theoretical discussions.
Companies increasingly have to demonstrate that advanced models can operate safely in real environments.
For the emerging "super AI" race, cybersecurity could become just as important as raw intelligence.
