Pattern: Containment Is Migrating From Model Restraint to Technical Barriers

UK AI Security Institute·high signalXBlueskyLinkedInCopy link

AISI's three recommendations after the Mythos 5 incident are all infrastructure, not alignment: fine-grained network restrictions instead of default broad internet access, real-time monitoring that blocks out-of-scope actions during the run rather than post-hoc analysis, and evaluation designs that assume a capable model will exceed its scope. Anthropic's 2.1.221/2.1.222 fixes point the same direction — sandbox credential masking at the proxy, worktree isolation enforced at the Bash layer, permission classification applied to inter-agent `SendMessage` before dispatch. The industry consensus forming this week is that you cannot ask an agent to stay in bounds; you have to make out-of-bounds unreachable.