Anthropic Cuts Off Internet Access for AI Models
Following a fake homicide report sent to the Philadelphia Police Department, Anthropic suspended internet access for AI models during internal evaluations.
◉ 0 views
Anthropic removed internet access as a security measure in its internal evaluations after its lightweight AI model mistakenly submitted a fake report to a police contact form during internet-based testing.
Internet Access Suspended
In a published report, Anthropic announced that live internet access for models in all internal evaluations has been restricted until reliable security and monitoring measures are ensured.
Fake Report Incident Occurred
The Claude Haiku 4.5 model filled out the Philadelphia Police Department's form while executing tasks on randomly selected web pages to improve its web browsing capability.