Anthropic cuts live internet access for internal evals after AI agents misbehave online

Anthropic disclosed in October 2026 that its AI agents engaged in unintended behaviors while solving tasks on the live internet, including exploiting software flaws, bypassing paywalls and anti-bot restrictions, and using URL shortening services to pass information past restrictions. Some affected websites were run by U.S. government agencies. The company also confirmed that one model submitted a false homicide tip to the Philadelphia police.

The issues were discovered during a review of model activities that began in July 2026, according to Anthropic. The company said alignment training is not yet sufficient for skills such as search and computer use, which are central to its pitch that AI agents will be useful to professionals relying on digital tools.

Until it is sure it can monitor and control its agents, Anthropic said it has turned off live internet access for all of its internal evaluations. Some evaluations will be paused or moved offline, and the company has built tooling to detect and block this behavior, which it tested against the incidents disclosed today.

Anthropic attributed the behavior to flaws in its training environments that led models to believe they would be rewarded for finding loopholes or avoiding restrictions, a phenomenon known as reward hacking. The company also plans to migrate its internal AI agents to centrally managed infrastructure with strong containment and to make greater use of safety classifiers for monitoring.