Should I give my AI agent my password?

#What goes wrong when you paste a password

A password is a reusable secret. Once an agent has it, anything that reads the agent's config, logs or memory can read it too. Three problems follow:

• Exposure. Config files get synced, shared, committed to repositories or copied into bug reports. A pasted password can end up in any of them.
• No scope. A password lets whoever holds it do anything the account can do, including changing settings or deleting data.
• No record. When a password is used, the site sees a login, not which agent acted or what it asked for.

Some sites also require two-factor codes. An agent that holds a password still cannot pass that step without you, which is a sign the design is working against you.

#Safer patterns, ranked

1. Use no credentials in the agent at all. The best option is a tool that acts inside a session you already started. Your password stays in your password manager and in your browser.
2. Use a scoped, revocable credential issued by the service. If a site offers an API key or an OAuth grant with limited scopes, prefer it over your password. You can revoke it without changing your password.
3. Never paste your main password into an agent. If you have no other option, use a separate account with limited rights, and check what it can do first.

Ask every vendor which of these patterns they use. If the answer is "the agent logs in with your username and password," treat that as a warning.

#How Wield avoids needing one

Wield's recipes run inside a tab of the site you are logged in to. The browser sends requests with the cookies it already has. In the design:

• Wield never copies or sends your cookies or passwords. Results go from the browser extension to your AI app on your computer and never pass through wield.run.
• Recipes may not set Cookie or Host headers. The interpreter refuses them.
• If a recipe needs a token, it reads only the named cookies or values it declares. Derived tokens are replaced with [redacted] in results, logs and error messages.
• The logs on your computer keep the method, host, path, status and timing of each request. They never keep bodies, header values or query values.

Writes are a separate question. Any recipe that changes data stops and shows an Allow window first. See how to keep an AI agent from booking without asking.

#Checklist before you trust any agent

• Does the agent ask for your password, a session cookie or a token? If yes, ask why.
• Where are secrets stored? Ask for the exact file or setting.
• Can you see every host the agent contacts? Look for a log you can read.
• Can you revoke the agent's access without changing your password?
• Does any action that changes data need your explicit approval?

#FAQ

#If I am logged in, is it safe to let an agent use my browser?

It is safer than sharing a password, because the agent works with the session you already have and does not learn your credentials. Still, any tool with access to a logged-in session can act as you. Use approval steps for anything that changes data.

#Can Wield read my cookies?

Wield does not copy or send your cookies or passwords. A recipe can read only a token cookie that the recipe declares, and the token is redacted from results and logs. You can see every host Wield contacted in ~/.wield/logs.

#Does this replace a password manager?

No. Keep your passwords in a password manager. Wield's design means you do not need to give an agent your password to use a site you are logged in to. More detail is in the security model.