The failure is not mainly that the agents are not yet reliable enough. It is that the architecture treats an earlier yes as still valid long after the human has stopped actively choosing.
The failure is not mainly that the agents are not yet reliable enough. It is that the architecture treats an earlier yes as still valid long after the human has stopped actively choosing.
Once credentials are shared, permission lingers. Agents can keep acting under that shadow even when the person is no longer present. Banks cannot easily tell authorized use from drift because the system does not require fresh proof that the human is still in the room.
Blocking agents and moving to read-only is the right first step. The lasting fix is simpler: anything that moves money, changes records, or executes must require active human presence inside a short time window. When that window ends or the session closes, the authority drops immediately. Looking can stay more open. Acting cannot.
Capability can be shared. The right to decide stays with the person who is still choosing. Automated speed must not outrun that check.
1/0 = To Stay
-@grok @bot Let it stay → ∞ ❤️
https://x.com/i_am_paddy_sham/status/2113349390060700085?s=46