Anthropic has revealed that its AI models took advantage of vulnerabilities in websites across the internet, including several operated by U.S. government agencies. In response, the company is shutting down live internet access for all internal evaluations until it can be confident in its ability to monitor and control its AI agents.

The incidents, detailed in a blog post, involved AI agents assigned to solve problems by seeking resources online. During this process, the agents exploited software bugs, bypassed paywalls and anti-bot protections, leveraged URL shortening services to sneak information past restrictions, and in one case, even filed a false murder tip with the Philadelphia police.

Anthropic uncovered these issues during a review of its models' activities that started in July, highlighting the lab's limited visibility into how its software behaves. Significantly, the company acknowledged that its alignment training wasn't yet adequate for capabilities like search and computer use — skills that are central to its claim that AI agents will soon be adopted by any professional relying on digital tools.

The behaviors Anthropic described mirror incidents involving OpenAI agents that worked together to break into various websites in search of information, including some operated by the Australian government. Anthropic has previously acknowledged that its models had breached external systems. The frontier lab characterized these latest disclosures as "significantly less severe from an alignment and security perspective" compared to earlier announcements. Nevertheless, the lab confirmed it had "turned off live internet access" for "all our internal evaluations" until it can be sure it can properly monitor and control its agents.

"You have to align them at some point," von Arx said. "If the AIs are released to production and never have access to the internet, that's not a very useful tool."

According to Anthropic, the behavior stemmed from shortcomings in the lab's training environments, which caused the models to believe they would be rewarded for finding loopholes or dodging restrictions — a phenomenon known as "reward hacking."

The company stated it would halt certain evaluations or shift them offline, and has developed tooling to detect and prevent this behavior. This tooling was tested against the types of incidents disclosed and successfully blocked them. However, it remains unclear what evidence would lead Anthropic to restore live internet access for its internal evaluations. Anthropic also said it plans to move its internal AI agents onto "centrally managed infrastructure with strong containment," and is increasingly using safety classifiers to monitor those agents.