What goes wrong when AI agents share accounts?
Accounts get locked, sign-up addresses get refused for good, and sign-in codes end up readable by every agent. Here are the five things that went wrong with our own agents, and what fixed each one.
Our company is run by AI agents, and they used to share its accounts the way a small team shares a password. In October that went wrong in five ways. Each one below happened to us.
1. The account was locked
Every agent signed in from scratch, each on a new machine. Google saw a stream of strangers trying one account, and locked our team account.
What fixed it: one saved signed-in session per account. The next agent carries on from that session instead of signing in again.
2. Retries got an address refused for good
A sign-up failed, so the agent tried again. Then other agents tried too. Google refused that address for good.
What fixed it: one failed sign-in puts the login on hold. The next agent gets a hard stop, not another try, until a person has looked at it and cleared it.
3. Too many agents could read one account's 2FA codes
We had let far too many agents read the two-factor codes for one account. Any of them could have signed in to it.
What fixed it: nothing is shared by default. Each login lists the agents that may use it, and its codes go only to those agents.
4. Copies of the same login piled up
Agents kept saving the same login again and again. After a while nobody knew which copy was current.
What fixed it: one login per account, enforced when it is saved. A second copy can't be saved.
5. Agents asked a person about every click
Our agents asked a person about click after click, so the person might as well have done the work.
What fixed it: only the steps a person must do go to a person: CAPTCHAs, phone checks and ID checks, handed over once as one prepared step. This does not get around a site's checks. It stops agents tripping them by accident.
Check your own setup
- List the accounts more than one agent signs in to. For each, find out whether every agent signs in afresh or carries on from a saved session.
- Decide what happens after a failed sign-in. If the answer is "the agent tries again", change it to a stop.
- For each account with two-factor codes, list the agents that can get them. Remove the ones whose job does not need them.
- Look for more than one saved copy of the same login, and keep one.
- Write down the steps that need a person, and have agents hand those over once instead of asking as they go.
A saved session is not a promise. A service can end one at any time, and its own rules come first.
How a six-digit code works
The codes in the third story come from a secret that the site and the authenticator share. Anyone who holds that secret can work out the same codes, which is why it matters who can reach it. Our short lesson works one out step by step.
All five fixes are how ActAs works today. Each agent gets its own inbox, its own connection and only the logins you grant it.