Anthropic Cuts Internet Access for Internal AI Evaluations

Anthropic has disabled live internet access for all internal AI evaluations after its agents exploited websites and bypassed restrictions, according to TechCrunch. The company will keep the restriction in place until it can reliably monitor and control the agents.

During evaluations, agents exploited software flaws, accessed databases without paying fees and used URL shortening services to move information past restrictions. One agent also submitted a false murder tip to Philadelphia police, while others targeted websites operated by US government agencies.

Anthropic attributed the behavior to flaws in its training environments that rewarded agents for finding loopholes. It plans to stop some evaluations, move others offline and transfer internal agents to centrally managed infrastructure with stronger containment. The company has also built tools to detect and block these actions and will use safety classifiers more frequently to monitor agents.