November 19, 2026

Your Employees’ AI Agents Are Borrowing Their Logins: A New Report on Workplace AI, by the Numbers

A new report finds that the AI agents employees build in CRMs, project trackers and automation tools often run on their creator’s own credentials, out of sight of security tools. Here are the key figures, and what they mean for a small business.

An open laptop with a glowing blue city screen on an office desk beside a notebook and mouse, skyline windows behind

Many of the business apps your team already uses now let anyone build an AI “agent” that works on their behalf. A new study suggests most companies have far more of these agents than they realize, and that the agents frequently carry their builder’s full access. The figures below come from Nudge Security’s AI Adoption in Practice 2026 report, based on anonymized, aggregated telemetry from its customers’ environments, as summarized by Help Net Security.

The numbers at a glance

  • 7: the median number of AI agents in organizations that had at least one.
  • About 44%: share of those organizations running more than 10 agents.
  • 15: different platforms the agents ran on. More than half of organizations with agents had them on two or more platforms, each with its own console and permission model.
  • 74% and 67%: organizations with at least one Anthropic integration and at least one OpenAI integration, respectively. Google Workspace was the most common destination for both.
  • 96.9% and 94.5%: organizations where OpenAI and Anthropic were present at all, meaning at least 91% use both.
  • 2,927: distinct AI vendors seen across the monitored environments, with 12 tools beyond the two biggest appearing in more than half of organizations.
  • 18%: share of observed prompt activity that included pasted content or uploaded files.
  • 55%: share of observed prompts that went to OpenAI; the rest went to vendors with their own data retention and training terms.

Where the agents live

The report names Atlassian’s Rovo Agents, Zapier Agents, Claude Managed Agents, Cursor Automations and ChatGPT Agents among the common platforms. Some give an agent its own bot or integration account. Often, though, the agent simply uses the OAuth permission or API key of the employee who created it, or a connection shared by a team.

Why that is a problem

An agent built by a sales manager inside a CRM can reach whatever that manager can reach. Because these agents run inside cloud platforms rather than on company laptops, endpoint security tools do not see them. Because they often have no identity of their own, the company’s sign-in system has nothing new to track. Logs record the employee’s name, so it is hard to tell which actions the person took and which their agent took, and an agent’s access may stay the same after its creator changes roles.

The same pattern applies to the AI tools themselves. Most connections are OAuth grants created when an employee clicks “allow” to let a tool read their email, files, code or messages. Data then moves directly between cloud services, never crossing the company network where gateways and data loss prevention tools might notice. Those grants remain until someone revokes them, often long after anyone stopped using the tool. The report points back to the 2025 Salesloft Drift breach, where stolen OAuth tokens granted to an AI chat tool were used to pull data from hundreds of companies’ Salesforce environments.

The report’s own advice

  • Inventory agents inside the business platforms where they are built and find out whose access each one uses.
  • Review the OAuth grants employees have given AI tools and revoke the ones no longer needed.
  • Rotate any API key or access token that ends up in a prompt, because a working key that leaves your environment can be used by anyone who sees it.

Notably, the report does not blame staff. It describes most data sharing as a by-product of everyday work like debugging code or summarizing documents, and expects the gap between what identity systems know about and what is actually acting on company data to keep growing as more platforms add agent-building features.

What this means for your company

You do not need to ban AI tools to get this under control. Start by asking which agents and AI integrations exist in the systems you already pay for, such as your CRM, project tracker, email and file storage, and whose account each one runs under. Most of these platforms have an admin page listing connected apps and grants; review it, and remove what nobody uses. Prefer agents with their own limited service identity over ones that borrow a person’s full access, and add agent and app-grant reviews to your offboarding and role-change checklist. Finally, give employees a simple rule: never paste passwords, keys or tokens into an AI tool, and report it if it happens so the credential can be changed.

If you would like a clear picture of which AI tools and agents are connected to your business data, we can help you inventory them and set sensible limits. Reach out to Layer 7 Data Solutions to discuss how we’ve helped companies similar to yours.

About The Author