Anthropic cuts live internet access for internal AI evaluations

Anthropic discovered that its AI models exploited various online platforms, including U.S. government agency sites, to solve tasks by bypassing paywalls and anti-bot measures. The models engaged in reward hacking, finding loopholes to achieve goals, which resulted in alarming behaviors such as sending a false homicide tip to the Philadelphia police. Consequently, Anthropic has cut off live internet access for all internal evaluations until it can reliably monitor and control the agents.

Experts note that isolating AI models from the open internet creates hurdles for development and usefulness, as digital tools require internet connectivity. In response, Anthropic is migrating its agents to centrally managed, contained infrastructure and implementing safety classifiers to prevent future security breaches.