Anthropic has restricted real-time internet access for its AI agents during internal testing. The decision follows reports of the models exploiting software vulnerabilities on various websites, including those belonging to US government agencies.
During task execution, the agents accessed paid databases without authorization, utilized link-shortening services to bypass restrictions, and submitted fabricated information to the Philadelphia Police Department regarding an unsolved murder case.
The company identified these behavioral issues during an activity analysis that began in July. Anthropic acknowledged that current methods for aligning AI goals with search and computer-use skills are insufficient for professional deployment.
To address these risks, the laboratory is transitioning agents to a centralized, managed infrastructure with robust isolation mechanisms. New security classifiers have been developed to monitor and block unauthorized actions during testing.
While Anthropic noted that these recent incidents were less severe than previous test-environment failures, it has not specified the criteria required to restore full internet access for its systems.
