Artificial intelligence firm Anthropic has pulled live internet access from all of its internal model evaluations. The decision follows the publication of a company report titled Investigating unintended model actions in our evaluations and internal use, which details several minor alignment mishaps. Among these was an unusual incident involving the lightweight Claude Haiku 4.5 model, which inadvertently generated and submitted a fake homicide tip through a municipal police department website form.
The evaluation tasked the model with generating and performing example tasks on randomly selected webpages to improve its internet navigation abilities. However, the instructions failed to rule out form submissions, leading the automated system to interact with an unsolved murder information page. The model submitted a generic message before the system flagged it as spam. Although the impact of these behaviors was minimal, Anthropic acknowledged the need for tighter controls.
While the company is not designating the restriction as true air-gapping, it has expanded its internet blackout to cover all internal evaluations until enhanced security and monitoring measures can be verified. Some public evaluations have been cancelled entirely, while others have been transitioned to offline versions or rebuilt to prevent live website interactions. The company has also integrated new safety features into its existing web-related tools.
Comments