SIAgentNews.com
← All news

Employee-built AI agents often run on their creator's login, and few teams track them

A Nudge Security report says agents built in everyday work apps borrow their maker's access, slipping past identity tools, device security and clear audit logs.

· First reported by Help Net Security

44%of organizations with agents had more than 10 of them

The short version

  • Organizations with agents had a median of seven; about 44% had more than 10, across 15 platforms.
  • Many agents use the builder's OAuth grant or API key, or a shared connection, not a dedicated account.
  • Logs show the employee, so a person's actions and their agent's actions are hard to separate.
  • 18% of observed prompt activity involved pasted content or uploaded files.
  • Nudge advises inventorying agents, finding whose access each uses, and revoking unneeded OAuth grants.

Why it matters

Tools your colleagues already use can now host software that acts with a person's full permissions, around the clock, without any new account being created. If your company's security checks only cover devices and official identities, these agents may sit outside them. The fix starts with simply knowing what exists.

Agents that act as their maker

As Help Net Security reported, a new study from Nudge Security finds that AI agents built by employees inside ordinary work software often operate with the credentials of whoever created them. The software in question includes project trackers, customer relationship management (CRM) systems and automation tools. The agent inherits whatever that person can see and do.

The report, called AI Adoption in Practice 2026, draws on anonymized, combined data from Nudge Security's customer environments. It tracks which AI tools appear, what they connect to, which agents are running and how staff use AI. Because the sample is Nudge's own customer base, it may not mirror every company.

Picture a sales manager who builds an agent in the company CRM. That agent can get to everything the manager can get to, whether or not the task needs it.

Why the usual controls miss them

Among organizations with at least one agent, the typical count was seven, and roughly 44% ran more than 10. Those agents were spread over 15 platforms, and over half of the agent-using organizations had them on at least two. Each platform has its own admin console and its own permission rules, so checking them means several separate reviews.

Examples named in the report include Zapier's agents, Cursor's automations, Atlassian's Rovo, and the agent products from Anthropic and OpenAI. Some platforms give an agent its own bot or integration account. Often, though, the agent borrows the creator's OAuth grant (a permission an employee approves so one app can reach another) or API key, or it uses a connection shared across a team.

That matters for two reasons. These agents run inside the platforms, not on company laptops, so endpoint security tools never see them. And if an agent has no identity of its own, the identity provider, the system that manages who has access, has nothing new to track.

A muddier audit trail

The logging problem may be the hardest to fix. When an agent works under a person's login, the records name the employee. The report says it is hard to tell which actions the person took and which their agent took.

Access can also outlive its reason. The report notes that an agent's permissions may stay the same when its creator moves to a different role. A person who changed jobs might leave behind an automation that still reaches data their new job doesn't need.

Nudge's CTO, Jaime Blasco, framed it as a visibility question. The issue, he said, is less about whether agents are allowed and more about whether security teams know they exist. He expects the problem to grow, since every platform is adding agent-building features, and he said the problem will get "harder, not easier."

OAuth grants and data in prompts

The report also looks at how AI tools link to company systems. Nudge saw at least one Anthropic integration in 74% of organizations and one from OpenAI in 67%, with Google Workspace the most common connection point for both. Anthropic's links tended to reach deeper into business systems, even though its adoption was a little lower.

Most of these links are OAuth grants, created when someone clicks "allow" so an AI tool can read their email, files, code or messages. The data then moves directly between cloud services and never crosses the corporate network. That keeps it out of view of AI gateways and data loss prevention tools. Grants remain active until someone revokes them, often long after the person stopped using the tool.

The report points to the 2025 Salesloft Drift breach as a warning. Attackers stole OAuth tokens that customers had granted to Drift's AI chat tool and used them to pull data from hundreds of companies' Salesforce environments.

Everyday prompts matter too. Pasted text or uploaded files appeared in 18% of observed prompt activity, and OpenAI received 55% of observed prompts. The riskiest content is API keys and access tokens, and Nudge advises rotating any that end up in a prompt. It doesn't blame employees, though. It says the sharing is mostly a side effect of AI being useful for jobs like debugging code or summarizing documents.

What security teams can do now

Nudge's advice is practical. First, inventory the agents inside the business platforms where they get built, and find out whose access each one uses. Second, review the OAuth grants employees have given AI tools and revoke those nobody needs.

Adoption data in the report also suggests why this is urgent. OpenAI showed up in 96.9% of organizations and Anthropic in 94.5%, so at least 91% use both. The report treats the major models as infrastructure, comparable to cloud providers that companies mix by use case. In all, Nudge counted 2,927 distinct AI vendors in the environments it monitors, and twelve tools appeared in more than half of organizations.

Our newsletter suggested earlier that giving agents their own accounts and narrow permissions is sensible general practice. That is our own advice, not a finding from the report, but it fits the problem the report describes: access that was never deliberately assigned to the agent.

What to watch

  • Whether platforms start giving agents their own dedicated identities by default instead of reusing a person's access.
  • Whether identity providers and security tools add ways to detect agents running inside business platforms.
  • Whether other security vendors' data show the same agent counts, since this report draws only on Nudge's customers.

Try it

List every agent your team has built in your CRM, project tracker or automation tool, and note whose login or API key each one uses. Then review which AI apps have OAuth access to your accounts and remove any you no longer use.

Sources and links

First reported by Help Net Security. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

More AI news