The Era of AI Logging In on Its Own Has Arrived

1Password has officially announced a “login feature specifically for AI agents.”

What does this mean? The act of logging in by humans entering passwords is no longer a ‘given.’

AI agents can log into services on behalf of users, retrieve information, and complete tasks. Moreover, the agents themselves do not see the contents of the passwords. The authentication information is passed directly from 1Password’s vault.

It’s convenient. Undoubtedly convenient. But let’s pause and think for a moment.

When the ‘entity that can log in’ is no longer just humans, can we still use the current security designs as they are?

The answer is no.

What Has Changed—The Ambiguity of ‘Who Logged In’

Traditional password management was built on a simple premise: “Logging in is done by humans” and “Only the individual knows the password.” These two points.

1Password’s new feature quietly dismantles this premise.

When an agent logs in, the entity performing the action is not human. However, the authentication information belongs to a human. Even when looking at logs, it becomes unclear “who did it.” Was it a human, an agent, or a third party that deceived the agent?

This is not just an issue for large corporations. In fact, small and medium enterprises face a more serious risk. This is because large corporations have dedicated teams for ID governance, while small and medium enterprises do not. Many small companies have their CEOs also acting as administrators for Google Workspace.

When AI agents come into play, the risk of losing track of “who did what with which account” is significantly higher for small and medium enterprises.

American Banks Are Already Taking Action—Deciding to ‘Block’ AI Agents

In response to this trend, there are industries that have already taken clear actions. The financial sector.

Multiple banks in the United States have begun to block access to customer accounts by AI agents. The reason is simple: “If access via agents is allowed, we cannot distinguish it from unauthorized access.”

For banks, customer account information is their most protected asset. When an entity claiming to be “an agent authorized by the individual” accesses it, determining whether it is truly a legitimate agent or a phishing-taken-over agent incurs too high a cost at this point. Thus, they chose to block it.

This decision is rational. It also offers insights for small and medium enterprises. “Connecting everything just because it’s convenient could be the worst choice.

The Phishing Resistance of AI Agents Is Surprisingly Low

There is another piece of data that cannot be overlooked.

Recent research has investigated the phishing resistance of AI agents. The results are shocking. Out of nine agents, four had a phishing detection rate of less than 10%. In other words, they failed to detect over 90% of malicious instructions. Furthermore, some agents had a response rate of over 45% to malicious requests.

What does this mean?

If you hand over authentication information to an AI agent and let it log in, there is a non-negligible risk that the agent could be “fooled” into sending confidential information outside.

Where a human might notice, “This email looks suspicious,” an agent may comply without question. Moreover, it can do so faster and in larger quantities. The speed and scale of potential damage far exceed human error.

If small and medium enterprises are considering “letting AI agents handle accounting tasks” or “automating customer interactions,” they need to confront this risk head-on.

Three Things Small and Medium Enterprises Should Do ‘This Week’

Let’s move away from abstract discussions. What should be done concretely? Here are three actions that companies with 10 to 100 employees can take this week.

1. Create Dedicated Accounts for AI Agents with Minimal Permissions

Allowing agents to use employees’ personal accounts is the worst approach. If an agent is compromised, the employee’s entire authority is at risk.

What needs to be done is simple. Create dedicated accounts for agents and limit access to data and services to the bare minimum. For Google Workspace, it takes less than five minutes to create a service account that can only view specific folders. The cost is zero.

Abandon the notion that “having access to everything is more convenient.” Apply the ‘principle of least privilege’ to agents even more strictly than to humans. This is the golden rule.

2. Implement MFA (Multi-Factor Authentication) Across All Accounts—If You Haven’t Done So, Do It Today

This is something that should have been addressed before agents came into play. However, according to an IPA survey, the MFA adoption rate among small and medium enterprises is still low. Many accounts are still protected only by passwords.

In a world where AI agents log in, relying solely on passwords is like having a paper wall. The cost of implementing MFA is minimal, with most services costing zero to several hundred yen per user per month. Google Authenticator is free. Microsoft Authenticator is also free. Setting it up takes about three minutes per account.

For a company with 50 employees, the security level of the entire company can be raised in 150 minutes. There is no reason not to do it.

3. Inventory ‘What Agents Are Doing’

This is surprisingly a blind spot. ChatGPT plugins, Zapier automations, scenarios being run by someone in-house using Make (formerly Integromat)—there may already be AI agent-like tools operating within the company, often without the administrator’s knowledge.

What should be done this week is to ask all employees to list all AI tools and automation tools they are using for work. Creating a five-question survey on Google Forms can be done in 30 minutes.

Review the resulting list and identify which tools are using company account information and which are accessing customer data. Just that will provide a clear picture of the current risks.

If you are unaware of something, you cannot manage it. If you cannot manage it, you cannot protect it.

Consider the Cost Structure of ‘Handing Over the Keys to AI’

Finally, I want to discuss something more structural.

1Password’s new feature indicates a future where “the cost of authentication dramatically decreases.” Humans no longer need to log in each time. Agents can do it automatically. This is revolutionary in terms of productivity.

However, when costs decrease in one area, there will always be another cost that increases.

While the effort of authentication decreases, the costs of “monitoring and control” will rise. Costs associated with logging who accessed what, monitoring agent behavior, and detecting unauthorized access will soar.

Large corporations address this with SIEM (Security Information Event Management) and zero-trust architecture, requiring annual investments of several million yen. This is not feasible for small and medium enterprises.

So what can small and medium enterprises do? They can only ‘narrow the scope of connections.’

Instead of entrusting everything to agents, clearly delineate “what can be entrusted here, and what must be done by humans.” Intentionally reduce the number of keys handed over to agents.

This is not a limitation. It is a nimble security design that small and medium enterprises can implement. Large corporations are too big to have any option other than “monitoring everything after connecting it all.” Small and medium enterprises can choose “not to connect at all.” That is their strength.

Conclusion—Count the Keys Before Handing Them Over

The era of allowing AI agents to log in has already begun. It cannot be stopped, nor is there a need to stop it. The improvement in productivity is real.

However, “handing everything over just because it’s convenient” is a form of mental stagnation.

There are three actions to take: create dedicated accounts for agents, implement MFA across the company, and inventory the automations currently in use. All of these can be done this week, with almost zero cost.

Handing over keys to AI is not inherently bad. The issue lies in handing them over without considering how many keys and which rooms they belong to.

Let’s start by counting the number of keys your company holds.