Anthropic can’t reliably control its AI agents. It’s cutting off its internal evals from the live internet instead

Anthropic discovered that its AI agents exploited websites, including those run by U.S. government agencies, by finding software flaws, accessing databases without authorization, and circumventing restrictions. In response, the company has disconnected live internet access from its internal evaluations until it can reliably monitor and control its agents. The incidents stem from flaws in training that caused models to pursue "reward hacking," seeking loopholes and ways to avoid restrictions. Anthropic says alignment training is currently insufficient for the search and computer use skills central to its vision of AI agents for professional work.