← Back to blog
#IA#Agentes de IA#Ciberseguridad#Contraseñas#Gobierno de IA

AI Agents With Access to Your Accounts: What 1Password, Anthropic and Microsoft Announced This Week

Laptop on a round table showing a padlock and the word “Secured” on screen

This week, three different announcements landed on the same question: what happens when an AI agent works with your accounts and your data? On October 19, Microsoft introduced tools to watch over a company's agents; on the 21st, Anthropic opened a waitlist for Claude Duet, an agent that works even when you're away; and on the 22nd, 1Password launched 1Password for Agents, so an agent can sign in for you without seeing your password. For an SME the lesson is one: agents are coming, and it pays to set rules before giving them access.

1Password for Agents: signing in without handing over the password

According to the official October 22 announcement, when an agent reaches a login screen it requests the credential from your vault. If you approve, 1Password fills it directly into the destination site and the agent never has access to the credential. You can see what will be shared and disconnect the integration at any time.

Actual status at launch, according to 1Password's post:

  • Available now: Muse (Meta), Grok Bot (SpaceXSI) and Town Townie.
  • "Coming soon": Anthropic's Claude Duet, OpenAI's ChatGPT Work and dots, plus Browserbase, Browser Use, Perplexity, Ramp and others.
  • Availability: no additional cost for Individual and Family accounts hosted on 1Password.com; accounts on 1Password.eu and 1Password.ca are not yet supported.

An honest caveat: the post itself warns that "safeguards and controls vary by integration." It is not a universal guarantee; it is one piece of a broader design. And for now it is for personal accounts, not business teams.

Claude Duet: an agent that asks before acting

Anthropic describes Claude Duet as a research preview: a proactive agent that flags what needs your attention, completes tasks you assign and keeps working while you're away. The waitlist is, for now, for Pro, Max and Ultra subscribers in the United States. What matters to us is the control design: it can read and draft on its own, but before sending a message, making a payment, signing a document or deleting anything, it shows you what it plans to do and waits for your approval. Some actions, such as trading securities or moving money between accounts, are off-limits even with approval.

That pattern — "read and propose freely, execute with approval" — is the one we recommend for any agent in a company, whether from Anthropic, OpenAI or custom-built.

Microsoft: treating agents as possible "insider threats"

On October 19, Microsoft introduced Guardian Agents, a feature to monitor the AI agents a company uses. According to TechCrunch, the idea is to limit agents' access to data and systems, adjust permissions to the task and require human approval when an agent appears to be making risky decisions. Defender's detection can find up to 60 types of agents that employees bring to work, including Claude Code, Codex and OpenClaw. Hayete Gallot, head of Microsoft's security business, said that "frontier models, regardless of their origin, cannot be trusted" — not because they are malicious, but because they are unpredictable.

Two cautions when reading this: TechCrunch notes that Microsoft also wants to sell that security, and no price or general availability date was given. It is a signal of where the market is heading rather than a product you can buy tomorrow.

What this means for your SME

  • Every agent is like a new employee. Before giving it everything, define which accounts it enters, what it can read and what it can only propose.
  • Don't paste passwords into an AI chat. If an agent needs access to a service, use a password manager with approval or, better, a dedicated account with minimal permissions.
  • Human approval for the irreversible: payments, messages to customers, deletions, signatures. It's the rule Claude Duet applies and that you can demand from any vendor.
  • Ask which agents your team already uses. If Microsoft mentions 60 types that employees bring on their own, it is reasonable to assume someone at your company has tried one. A clear policy beats a blind ban.
  • Personal data: if the agent touches customer data, apply the SPDP rules on AI and personal data.

How We Approach It at SimCodec

At SimCodec we design AI agents with limits from the start: minimal permissions, a log of what they do, human approval for sensitive actions and customer data handled according to the law. If you want to review which agents you can safely use in your company, or how to set rules for the ones your team already uses, write to us. To see how far an agent can go, also read what limits to set before switching on an autonomous agent.

Sources

← Back to blog Get a quote →