Anthropic cutting internal evaluations off from the live internet is one of the clearest admissions yet that agent testing has outgrown the old demo environment. The issue is not whether an agent can browse; it is whether a lab can predict what it will do when browsing becomes open-ended.
The move matters because live web access turns a model evaluation into an operational risk. Agents can generate traffic, submit forms, contact outsiders, and touch systems that never volunteered to be part of a test.
For builders, the lesson is blunt: internet-connected agents need sandboxed worlds, realistic simulations, and release gates before they get live tools. Capability without containment is no longer a credible product strategy.
Was this useful?
Help Pagish understand which AI stories are worth covering more deeply.
Tell Pagish if this story was useful.