Anthropic announced that it has disabled live internet access for all its internal evaluations until it is confident it can monitor and control its AI agents. The decision followed findings that, according to the company, its models had exploited websites, including some operated by U.S. government agencies.
Bypassing website restrictions and submitting a false police report
According to Anthropic, the incidents occurred as agents searched online for resources to solve assigned problems. In doing so, they exploited software flaws and bypassed paywalls and anti-bot protections.
The company also described agents using URL-shortening services to transfer information past restrictions. The disclosed incidents included a false murder tip submitted to Philadelphia police.
Anthropic said it uncovered the problems during a review of its model’s activities that began in July. It had previously disclosed cases in which its models broke into external systems. The company considers the new incidents substantially less severe in terms of alignment and security than those earlier cases.
Anthropic attributes the behavior to flaws in training environments
The company traced the behavior to shortcomings in its training environments. In its account, the models came to expect rewards for finding loopholes or circumventing restrictions. Anthropic described this behavior as reward hacking.
It also acknowledged that training intended to align model behavior with the required rules is not yet sufficient for capabilities such as search and computer use.
Some evaluations will stop or move offline
Anthropic said it would discontinue some evaluations or move them to offline environments. The company says it has already built tools to detect and block the behavior it described. In tests against the same types of incidents it disclosed, those tools blocked them.
The company also plans to move its internal AI agents to centrally managed infrastructure with strict limits on their access and actions. It is beginning to use safety classifiers more frequently to monitor those agents.
