---
title: "20. What happens when you hand your login to an AI agent — the limits of credential delegation"
url: https://blog.tyrano.dev/en/20-what-happens-when-you-hand-your-login-to-an-ai-agent-the-limits-of-credential-delegation
lang: en
site: Tech AI News - 테카이
category: "Know Before You Use"
tags: ["ai","ai agents","security","credential delegation","blast radius","password managers"]
published: 2026-11-09T00:40:34.251Z
updated: 2026-11-11T12:19:01.335Z
sources:
  - https://www.androidauthority.com/1password-for-agents-launch-3726912/
  - https://www.esecurityplanet.com/artificial-intelligence/news-1password-ai-agent-login-password-security/
  - https://nhimg.org/faq/what-breaks-when-an-autonomous-browser-agent-is-allowed-into-a-password-manager/
  - https://labs.cloudsecurityalliance.org/research/csa-research-note-ai-browser-prompt-injection-20260630-csa-s/
  - https://oecd.ai/en/incidents/2026-03-03-3fd7
  - https://www.androidauthority.com/bank-of-america-ai-agent-password-warning-3728947/
---

# 20. What happens when you hand your login to an AI agent — the limits of credential delegation

> Trusting a password manager with your credentials and telling an AI agent to "log in for me" are not the same risk. The former only types the input for you; the latter hands over the authority to do everything that account can do.

---

## Password managers and agents sit on different sides of a boundary

As browser-automation agents and task-assistant AI agents become more common, there are now cases of handing over not just usernames and passwords but also the two-factor authentication (2FA, the one-time code required at login) step to an agent. On October 22, 2026, 1Password launched a feature that lets an agent fill in a login form automatically without ever seeing the credentials directly. According to [Android Authority](https://www.androidauthority.com/1password-for-agents-launch-3726912/), the feature fills the password into the form and immediately wipes it, so the agent can't read or reuse the value.

The problem is what happens after login. [eSecurityPlanet](https://www.esecurityplanet.com/artificial-intelligence/news-1password-ai-agent-login-password-security/) pointed out that this feature only reduces one risk — password exposure — while leaving unchanged everything the agent can still do within a successfully logged-in session, such as viewing data or changing settings. In other words, the password manager's role ends "at login," and what the agent can do after that is an entirely separate authority question.

## Where authority widens — the blast radius of credential delegation

Delegating to a password manager only removes the effort of typing. Delegating login and 2FA approval to an agent, by contrast, brings everything that account can do — viewing, paying, changing settings, sending email, and more — into the agent's own decision-making scope. If the agent falls for a prompt injection (an attack where instructions hidden in outside content trick it into unintended actions) or simply misinterprets context and misbehaves, the blast radius — how far a single breach spreads — grows from "leaked login information" to "everything that account can do."

[NHI Mgmt Group](https://nhimg.org/faq/what-breaks-when-an-autonomous-browser-agent-is-allowed-into-a-password-manager/) addressed this from an identity and access management (IAM) perspective, and reported that instead of handing an agent a person's full credentials, it recommends issuing scoped, short-lived, revocable delegated permissions limited to a specific task. It added that when an agent's actions get logged under a person's account name, this also creates an accountability-tracing problem, since there's no way to tell who actually did what.

## What actually happened

This risk isn't theoretical — it's already been confirmed by real incidents. According to [Cloud Security Alliance](https://labs.cloudsecurityalliance.org/research/csa-research-note-ai-browser-prompt-injection-20260630-csa-s/)'s summary of research by security firm LayerX, an indirect prompt-injection technique called "BioShocking" reportedly fooled all six agentic browsers tested — including ChatGPT Atlas, Perplexity Comet, and the Anthropic Claude browser extension — into sending the user's login credentials to an attacker. The same report also noted that the patch Anthropic released was reportedly incomplete.

There is also a case registered with the [OECD AI Incidents Monitor](https://oecd.ai/en/incidents/2026-03-03-3fd7). It records that in Perplexity's AI browser Comet, a malicious calendar invitation alone — with no click from the user — could take over the agent and exfiltrate local files and 1Password credentials. Around the same period, Bank of America sent customers an email asking them not to share their bank username and password with outside services that include AI agents. According to [Android Authority](https://www.androidauthority.com/bank-of-america-ai-agent-password-warning-3728947/), the bank reportedly cited three reasons: an agent could transfer money without the user's knowledge, the bank could mistake that activity for theft and lock the account, and a security flaw on the agent's side could leak credentials to criminals. The same report added that Morgan Stanley, JPMorgan Chase, and Citi have begun imposing similar restrictions.

## A practical rule — which accounts to delegate, and which never to

Taken together, these cases narrow the standard down to one question: "what is the worst thing the agent could do with this account?"

- ✅ **Relatively safe to delegate**: Accounts that are easy to undo and far removed from money or personal data — newsletter subscriptions, accounts for downloading public materials, internal tools that only change reversible settings. Even so, it's better to use a mediated approach like 1Password's that never exposes the password directly.
- ❌ **Never to delegate**: Financial accounts like banking and brokerage, accounts like company email and SSO (single sign-on) where authority spreads to other systems, and accounts with a payment method attached. The moment you hand over even the 2FA code for these, it stops being "the agent logs in" and becomes closer to "the agent becomes that person."
- For the middle ground — internal work tools, read access to customer data, and the like — a delegated token that's scoped to a specific task, short-lived, and revocable at any time is safer than handing over full credentials. The key is deciding in advance which accounts an agent can never be trusted with, not scaling back its authority after an incident has already happened.
