AI

Risks of agents with tools, least-privilege design, sandboxing, human-in-the-loop approvals, budgets and timeouts, audit logs, testing for misbehaviour and incident response.