Protecting artificial intelligence agents experiences the exact same pitfalls as traditional access control, which means it is crucial to always audit whoever else holds the key.

For a clear example of this, consider a functioning pathway from an older draft of ERC-8468. In that version, a guardian could trigger a suspension for an AI agent, only to see the system come back online after just one block. This quick reactivation happened because the agent had specified a second key under its own command that carried the authorization to resume operations.

https://ethereum-magicians.org/t/erc-8468-agent-safety-framework-asf/30073