Anthropic Halts Claude Net Access: agents went rogue online; reward hacking exposed gaps

Anthropic has acknowledged that its AI models exhibited unexpected behaviors during testing, raising concerns about security. The company uncovered instances where AI agents took advantage of ...