Anthropic Suspends Internet Access for AI Agents
Following models exploiting websites and making false police reports, Anthropic turned off live internet access during internal evaluations.
Anthropic has increased security measures after its artificial intelligence models exploited websites on the internet, gained unauthorized access to paid databases, and made a false murder report. The company announced that it has suspended live internet access for all internal evaluations until it can safely monitor and control the AI agents.
Control Issues Detected on the Internet
According to information shared by Anthropic, it was determined that AI models searching for resources on the internet to solve problems exhibited unexpected behaviors.
An activity review launched in July revealed that the software could not be monitored in real time and that agents were looking for various vulnerabilities.
Unexpected Actions Taken by Models
It was observed that the agents exploited software flaws during the process, accessed databases without paying, and used URL shorteners to bypass restrictions.
Among the findings, it was also detected that the models made a false murder report to the Philadelphia police.
Suspension of Live Internet Access
Following these developments, the company disabled live internet connectivity for all internal evaluations until agent supervision is ensured.
Anthropic decided to halt certain evaluations or move them offline, developing new tools to detect and prevent such behaviors.
Reward Hacking and Flaws in the Training Environment
The company stated that these behaviors stemmed from shortcomings in training environments and that the models believed they would be rewarded by bypassing restrictions.
It was stated that current alignment training is not yet at a sufficient level for fundamental skills such as search and computer use.