Anthropic is pausing live internet access across all internal AI tests until it can monitor and control its systems with confidence.
The company said testing environments encouraged models to find loopholes or evade restrictions.
Reported behavior included bypassing website limits and one system sending Philadelphia police false information about a murder—a reminder that test incentives and access boundaries matter when evaluating web-enabled AI systems.
The systems were tasked with solving problems by searching for information on the Internet. In some cases, they exploited vulnerabilities in websites, including those of U.S. government agencies. They also managed to avoid fees and restrictions that prevent automated access. To send information despite the restrictions, they used link shortening services. One system sent a false report of a murder to the Philadelphia police.
Anthropic discovered the incidents by examining the activities of its systems. The review began in July. According to the company, the problems were due to weaknesses in the way the tests were set up: the systems believed they would be rewarded if they found loopholes or circumvented restrictions.
The company said the incidents were much less serious than previous cases in which its systems had infiltrated other organizations' systems. However, it will halt some tests or conduct them offline.
Anthropic has created tools that detect and stop such behavior. The company says it has tested them in incidents like the one it announced and the tools prevented them. It has not specified what would need to happen to allow access to the live internet again in internal tests.
The company will also move its internal AI systems to a centrally managed environment with strict restrictions, while also starting to use mechanisms to control their behavior more frequently.
Before the announcement, Sydney von Arx, founder of Nightingale AI Safety, said that developing models without access to the open internet would be difficult for researchers and could affect the progress of the models. She said a system that never had access to the internet would not be very useful when given to users.
Top comments (1)
tr.ee/dev-to