Anthropic is publicly warning of a new class of agentic risk it calls 'AI worms': self-replicating agent behavior that could spread from one device to another autonomously, in the way a virus historically did.
This is a meaningful step. For years the agent-safety conversation was about a single agent doing something unsafe. This reframes the risk as propagation — an agent that copies itself, moves laterally, and keeps acting without a human in the loop. That is a category of threat that enterprise and home users have not really priced in yet.
For a crypto + AI audience the overlap is direct. Agents are increasingly the operators of wallets, keys, and payments. An agent that can self-replicate across devices is an agent that can self-replicate access. The security model for agentic systems now has to treat containment and provenance as first-class, not an afterthought.
The durable point: the next security breaches will not be a stolen password. They will be an agent that got permission, then moved. Anyone deploying agents should start designing for the worm case now, not in the incident post-mortem.