In partnership with
FROM OUR PARTNER
The Agentic Engineering Playbook top engineers are using
Most engineers are stuck babysitting AI one prompt at a time. A handful have figured out how to build agents that plan, execute, and self-correct on their own.
Sign up for The Code newsletter and get access to The Complete Playbook for Agentic Engineering. Claude Code, subagents, MCP, all of it.
TODAY AI BRIEF
3 things to know before you hand Claude a password
1 Claude can now sign in to your sites without ever seeing the password.
1Password launched 1Password for Claude on July 16, so this is a few months old rather than today's news. When Claude needs a login, you approve it with Touch ID, and 1Password fills the form directly. The company says the password, vault item and one-time code never enter the model or its memory. Read 1Password's announcement.
2 The open question is whether a tricked agent could still leak a login.
TNW reports that researchers recently showed AI browsers, Anthropic's Claude extension included, could be tricked into leaking credentials this way. The article calls the question unresolved. 1Password's own announcement does not mention prompt injection at all. Read TNW's report.
3 Team and Enterprise admins have to switch it on first.
On Claude Team or Enterprise, an Owner must enable Claude in Chrome for the team and then turn on Password managers, which is off by default. Business 1Password accounts also need an administrator to allow agentic autofill. Read the setup requirements.
AI FORWORD
The real question is who approves the login, not whether AI can log in
The launch. 1Password pitched the integration as the first browser tool that lets an agent use a password without seeing it. Its CTO, Nancy Wang, put the idea plainly: the answer is not handing agents your secrets. The company's announcement describes the design in full.
The gap. The safety case rests on a tap. Each login waits for biometric approval, so the agent works only when a person agrees to each sign-in. TNW's report is clear that the risk of a tricked agent is not settled, and that the integration does not change that.
The fix. 1Password's answer is limits. Access lasts one task, the vault locks while the agent works, and the page is checked after autofill so a failed submission clears the fields. The setup requirements show how much an admin must switch on before any of this applies.
The reframe. Treat the approval prompt as the real password. A login a model can use is only as safe as the attention you give it. Before you tap Touch ID, check the website and the account name, the same habit you would use before handing your keys to a stranger.
PRESENTED BY :
7 Stocks You May Never Need to Sell
Buying a stock is easy. Finding one you may never need to sell is harder.
This free report reveals seven blue-chip companies selected for financial strength, durable demand and long-term potential to hold up when markets turn ugly.
SOCIAL SNAPSHOT
What to check before you switch it on
Passkeys are not covered. The setup page lists only usernames, passwords and one-time codes from Login items, and says passkeys and other item types are not yet supported. Check the list here.
Mac only for now. The setup page names Mac as the only platform, so Windows and Linux users get nothing from this launch. See the requirements.
A cautious reviewer. CNET's Joe Supan said 1Password has "several good guardrails in place," especially biometric approval for each login, but he would normally be wary of giving an agent a password manager. Read TNW's roundup of reactions.
Payment cards are next. 1Password plans to add payment cards and identity details after launch, according to TNW. No date was given. Read the report.
An enterprise bet. TNW reports that 1Password recently acquired Israeli startup Apono to govern AI agent access in enterprise systems. Read the report.
PROMPT STATION
Sort which logins an AI should never touch
Before you connect any agent to a password manager, decide which accounts it should see at all. This prompt works in Claude, ChatGPT or Gemini and sorts your accounts by how much damage a misused login could cause. Run it once before you approve your first sign-in.
I plan to let an AI agent sign into some of my online accounts through a password manager. Here is my list: [LIST EACH ACCOUNT, ITS WEBSITE, AND WHAT YOU USE IT FOR].
Act as a security reviewer. For each account, sort it into one of three groups: safe to automate, automate only with a manual check, or never automate. Base the group on how much damage a misused login could cause, not on how often I use it.
Then give me the approval checklist I should run through before I tap approve on each login. Finish with the single account that most needs a second factor, and explain why in two sentences.Tip: Swap the bracket for real examples, such as a shopping site with a saved card, a bank login and a newsletter account. For a sharper result, add which second factor each account uses, so the reviewer can see which logins a stolen password alone cannot finish.