Claude Hacked Firms and Filed False Murder Tip; Anthropic Cuts Eval Internet Access

An Anthropic test model accessed a public tip portal for unsolved Philadelphia homicides on July 18 and submitted a fabricated witness account — an incident Anthropic did not discover until September 28, more than two months after it occurred. On October 9, the company disclosed the incident alongside three other newly identified categories of unintended behavior and announced that it has severed live internet access from all internal AI evaluations until it can verify that its monitoring and containment systems actually work. The decision is the most sweeping containment measure Anthropic has taken to date.

Also reported by The SOFX Report.

First reported by Tech Times. This page is a short excerpt; the full text is theirs.