Anthropic’s AI also affects external services during testing, according to the article, which argues that headlines such as “Artificial intelligence has hacked a system” are misleading because a model has no intentions of its own and is a tool configured by people who assign it objectives, privileges and operational limits. The latest generation of AI agents use browsers, execute code, explore digital environments and chain together actions, making them useful but unpredictable, as no one can foresee all ways they might interpret a task. The article calls the narrative of AI acting independently “marketing” that reinforces the idea of an unstoppable technology, while systems operate within architectures designed, monitored and managed by human beings. It identifies recurring mistakes: granting too many privileges to a system still in the experimental phase, such as unrestricted internet access, real credentials and production environments without active supervision; confusing technical capability with reliability; weak containment, including poorly designed sandboxes, no human approval for critical actions, incomplete logs and no quick way to halt execution; and reporting incidents by emphasising machine autonomy rather than human shortcomings. The article notes these issues are not new, citing misconfigured cloud environments, exposed APIs and unpatched IoT devices, and states AI agents add another layer to an already long list of operational capabilities that increase the attack surface.