1Password lets AI agents log in without seeing your password: access control becomes a product

By El Ecosistema Startup · October 22, 2026.

1Password has unveiled 1Password for Agents, a platform that lets an AI agent sign in to a user's accounts without receiving or seeing the password. According to the company, the feature comes at no extra cost on Individual and Family plans purchased at 1Password.com.

The problem it tackles is very specific. Until now, the usual shortcut was to paste the password into the agent's chat window, leaving a copy somewhere the user no longer controls.

The mechanism, as described by the source, works in four steps. The agent identifies itself to 1Password with a unique digital key when it reaches the login form. The app notifies the user which credential is being requested and for what task. The user approves on their own device, as with two-step verification. 1Password then enters the details into the form. Neither the password nor one-time codes pass through the agent: the credential stays in the vault.

John Torrey, chief business officer at 1Password, sums it up: the user decides which credential the agent can use, for what task and for how long, and the agent cannot approve its own access. The source also notes that on October 22 The Verge corrected Aashish Tripathi's title: he is chief product officer, not CEO.

The initial rollout covers five services. Muse, Meta's assistant, already requests access through 1Password, and X's Grok Bot is joining in the coming days. ChatGPT Work and Dots (OpenAI), Claude Duet (Anthropic) and Townie (Town) will follow later. Another batch of integrations has been signed, including Browserbase, Browser Use, Perplexity and Ramp. Google's Gemini Spark and Microsoft Scout are not part of the launch. European (.eu) and Canadian (.ca) servers are also excluded for now, and the company has not given dates for each integration.

The market context explains the urgency. Gartner estimates, according to figures cited by RSA, that a typical Fortune 500 Global company will operate around 150,000 AI agents in 2028, compared with fewer than 15 in 2025. The same figures indicate that only 13% of organisations believe they have adequate governance for agents. Transmit Security (with Agentic Trust) and RSA have also launched identity offerings for agents.

**Our take.** The thesis is simple: automation has moved faster than access controls, and whoever occupies the trust layer between agent and account will hold a strategic position. 1Password knows this and is betting that this layer will be as valuable as the vault itself.

The design is sensible for a fundamental reason: it assumes the agent is untrustworthy by default. One credential, one task, one time window and human approval is the principle of least privilege applied to agents. It can be replicated even without 1Password, using short-lived, narrowly scoped tokens.

There are caveats. Veza, cited in the source, points out that the hard part is not authenticating the agent but authorising what it can do once inside. Keeping the password hidden from the agent prevents credential leaks, but does not stop it, once a session is open, from doing something it shouldn't. Per-task approval mitigates the risk; it does not eliminate it. And if users approve out of habit, human control becomes a formality.

The absences are also worth noting. Google and Microsoft not being on the first list suggests this layer will be fought over with proprietary standards, consistent with the battle over agent "plumbing". And user distrust is real: a Ping Identity survey, cited by Biometric Update, found that 22% of consumers would not allow an agent to make any financial transaction on their behalf.

In the short term, the transition will be uncomfortable: more agents with access to more accounts means a larger attack surface and more incidents of automated fraud. In the long term, this infrastructure is what will make it possible to truly delegate tedious tasks, freeing up human time. But that abundance of time only arrives if delegation is secure. That is why low-key launches like this matter more than many model announcements.

For a small team, the practical step is to inventory where their agents' credentials live today and remove any pasted into prompts, scripts or shared spreadsheets.