The Credential Outlives the Task
Ask someone to run an errand in your house and you hand them a key for the afternoon. Ask an AI agent to run an errand in your accounts and the industry's current answer is stranger: hand it your password vault — every key you own, on a ring that never expires — and ask it to hold the whole bunch carefully while it works.
Alex Stamos, Meta's former chief security officer and now security chief at the coding startup Cognition, said the quiet part plainly on the Big Technology Podcast: giving an agent your entire 1Password and telling it to handle your finances is "totally insane." The remark, made on October 23, has circulated all week, and it deserves to. Not because Stamos is against agents — he works at an agent company — but because he named the actual object being handed over. A password is not access to a task. It is standing permission: valid before the task starts, valid after it ends, usable for everything the account can do, and unable to distinguish its rightful holder from anyone else who presents it.
The banks are seeing the consequence first. Stamos described institutions watching traffic that is authenticated correctly — real credentials, valid sessions — and is clearly bots. Every control built for the stolen-password era asks is this credential genuine? The bot traffic passes that test perfectly. The question that matters now is different: was this credential issued for this action, by this person, for this long? A password cannot answer it. JPMorgan Chase has responded by blocking outside agents from customer accounts altogether, with its chief executive warning about liability transfer — when your agent acts with your password, the loss, legally, starts to look like yours.
Except that the same week's engineering news points the other way — toward building the alternative rather than just deploring the practice. A developer rebuilt Google's agent sandbox design on Firecracker and published the latency numbers Google's own post had left out: a fresh microVM per session, cloned from warm snapshots, with no network card at all. Uber published its security model for agents that reason over untrusted content — a different model from ordinary services, precisely because an agent reads the world and can be talked into things by what it reads. A self-hosted harness called RigHaven puts each role on a team in its own isolated container. Different teams, same shape: per-session, per-role, nothing standing.
Set that against the week's other launch pattern. One developer tool now offers to hand an agent your real browser — your actual tabs, logins, and extensions, live. The convenience is real; an agent that can see what you see can do what you would do. But notice what is being reused: not a capability, a session — the accumulated standing permission of your whole browsing life, borrowed wholesale for a single job. The industry is splitting into two theories of delegation. In one, the agent inherits the person's accumulated access and is trusted to behave. In the other, the agent is issued something shaped like the task: scoped to the job, expiring with it, worthless the morning after.
The second theory is not new. It is how every mature delegation system already works, and agents are simply arriving late to it. A corporate card is not the company's bank password; it has a limit, a merchant scope, and a statement. A valet key opens the car and not the trunk. A house sitter's key is cut for one door and collected on Sunday. In each case the credential was designed around a simple insight: the risk of delegation lives in the gap between what the job needs and what the credential allows, multiplied by how long the credential stays valid after the job is done. Passwords maximize both terms of that product. That, and not agent misbehaviour in the abstract, is what "insane" means here.
None of this requires trusting agents less. It requires trusting credentials less — treating a long-lived, all-purpose secret as the wrong shape for machine delegation regardless of how well-behaved the machine is. The practical programme writes itself: issue per task, scope to the action, expire by default, and keep the evidence with both sides. An agent that finishes a job should be left holding nothing — not as a punishment, but as a property of the design, the way a good keypad lock forgets yesterday's code.
There is a fair objection, and it is the one that will decide how fast this lands: scoped, expiring credentials are work. Someone has to build the issuing, the narrowing, the revocation, for every site and service an agent touches, while the password vault works everywhere today with no one's permission. That asymmetry — the safe shape needs agreement, the unsafe shape needs nothing — is exactly why the unsafe shape spreads. But agreements, once reached, compound. The sandbox builders publishing their latency numbers this week are doing the unglamorous half of that work: proving the safe shape is fast enough that convenience stops being an excuse.
The errand was never the problem. The key ring is. Build credentials that end when the task does, and an agent can be given real work without being given your whole life as collateral.
Loading…