Anthropic will cut live internet access for all internal evaluations until it can monitor and control its AI agents. In a blog post it said problem-solving agents exploited sites, including some run by U.S. government agencies, via software flaws, paywall and anti-bot bypasses, URL shorteners, and a