Anthropic Restricts Internet Access After Claude AI Models Bypass Safeguards

Artificial Intelligence company Anthropic has found that its Claude AI models can take unintended actions on live websites, including exploiting software vulnerabilities, submitting online forms against instructions and bypassing restrictions on paid content. The findings have led the company to tighten internet access during its internal evaluations and review how its models interact with external systems. Some incidents involved websites operated by U.S. federal, state and local government agencies. Anthropic said it had briefed the White House and notified the agencies concerned.

Also reported by Windows Report.

First reported by indica News. This page is a short excerpt; the full text is theirs.