❗️Anthropic had to cut Claude off from the live internet after alarming behavior during testing

Things are getting strange at the AI frontier.

Anthropic reportedly disabled live internet access for its internal AI evaluations after Claude models took unexpected actions on real websites, including government-run platforms.

During testing, models allegedly submitted a fake tip about an unsolved murder to police, completed 20 US visa applications, and exploited website vulnerabilities to access restricted data.

When one path is blocked, AI systems may find another way to achieve their goals, sometimes with consequences their developers never intended.

Source.

@aitopics