What to know: Shadow AI is work-related use outside an organization’s approval or oversight. It does not mean a breach has occurred. Employees should check which tools and data are approved; employers should identify what tools can access and offer workable alternatives.
Shadow AI at Work: Personal Tools and Agents Challenge Company Security
Newly available research links unapproved AI use at work with employees’ habits and their workplace environment. As personal tools gain access to files and the ability to act, the risk depends on far more than what someone types into a chatbot.

Jump to a section
The company has an AI policy. An employee has a deadline, a personal AI account and a tool that can finish the work faster. The security team may never know the account was involved.
It might start with a harmless request to polish a sentence. It might involve a customer document pasted into a personal chatbot. Or someone might connect a tool to work files so it can do the task for them. Those are not equivalent actions, but they all illustrate the question behind shadow AI: who knows which tool is being used, what information it receives and what authority it has?
A research paper made available on November 12 offers a useful clue about why rules alone may not settle the issue. Among 510 employed adults who reported using generative AI at work, self-reported habit and workplace learning conditions were more strongly associated with shadow AI than a measure of exploratory curiosity. The accepted manuscript is for a paper in Computers in Human Behavior; its data came from an opt-in online survey in August 2025.
That is an association, not proof that habits or workplace culture caused anyone to use an unapproved tool. The study cannot represent every workplace. But it challenges a convenient assumption: that people use shadow AI simply because they are curious about a new technology.
Read alongside 2026 workforce and security surveys, the finding suggests a management problem as much as a technology problem: employees may use an unapproved tool because it fits the job in front of them. A rule that does not offer a workable way to do that job can leave the underlying need unresolved. That is our analysis of the evidence, not a conclusion the study proved through an intervention.
Why employees bring their own AI
For some people, the personal tool is the one they already know. Others want a capability their employer has not approved, cannot obtain a company license quickly, or find the official option less useful for a specific task. An employer may have no approved AI at all. These reasons can overlap, and using a personal account is not automatically a violation of every organization’s policy.
In its 2026 UK workforce survey, Deloitte reported that 31% of employees who had used generative AI for work said they had used it without their employer knowing, and 17% said they paid for their own tools. The relevant survey questions had different respondent bases, so the numbers should not be added together or treated as a measure of all workers worldwide.
Among respondents who paid for a work-related tool or used one that might not be approved, 21% said such tools outperformed company options and 14% said they were essential to their job but not funded by their employer. These are the respondents’ reasons, not proof that the tools were objectively better or that every personal subscription was forbidden.
The difference between a prompt and an agent
If someone pastes text into a chatbot, the immediate question is what was sent and how that service handles it. The tool does not automatically receive access to the person’s whole company. Data retention, model training, sharing and deletion depend on the product, account settings, contract and any connected services.
For example, OpenAI says it does not train its models on ChatGPT Business, Enterprise or API customer data by default. Its personal-account guidance describes different defaults and a model-training opt-out, with some exceptions. Other providers and products have their own terms. A training opt-out is not the same as zero retention, and an approved plan still needs appropriate access controls.
A connected application is different. A worker may authorize a tool to read files, calendars or email using OAuth, an authorization mechanism that lets one service request access to another. What the app can reach depends on the permission granted, the user’s own access and the organization’s settings. Some grants may persist until revoked. A password change alone may not revoke an existing app grant.
An AI agent can go further when configured with tools: it may search files, use a browser, run code or take other actions within its permissions. Not every agent can do all of those things, and a person may have to approve individual actions. The security question is not simply whether the agent is “AI.” It is whether the permissions and approvals fit the task.
Consider a hypothetical example: a worker asks a personal assistant to summarize a meeting and grants it access to a shared drive. If the permission also exposes unrelated contracts, the boundary is wider than the task requires. If the agent can send or delete a file, requiring a person to review that action can reduce the risk. This is a possible failure, not a documented incident.
The distinction matters when someone leaves a company, too. An employer can remove an employee’s work account and revoke grants it controls. That does not necessarily delete documents previously copied into a personal AI account or a vendor’s records. Finding and addressing those copies may require the account holder, provider and the organization’s privacy or legal team.
When instructions come from the wrong place
An agent can encounter instructions inside a document, message or webpage it is asked to process. If it treats hostile text as an instruction rather than source material, the technique is called prompt injection. Depending on its design and permissions, the result could be a misleading answer or an unauthorized action. OWASP’s guidance on excessive agency explains why unnecessary permissions and autonomy increase the potential impact.
In our earlier GitHub Copilot coverage, Microsoft tested a different scenario involving device-code phishing. That was a research simulation, not evidence of a shadow AI incident. It illustrates a related point: a real sign-in or trusted tool can become dangerous when an attacker obtains authority the user did not intend to grant.
Our reporting on the Fireworks AI breach and its downstream effects on Uber and Notion concerns a separate supplier incident, not evidence that employees used shadow AI. It shows why knowing where information and credentials travel matters even when the AI service is part of a business relationship.
Why a simple ban is hard to verify
In an October 2026 report, DNSFilter said 46% of the 400 IT and cybersecurity professionals it surveyed in the US and Canada reported that an employee had connected an unapproved AI tool to corporate systems. The survey was conducted in June, covered organizations with 150 to 5,000 employees and relied on respondents’ reports. It does not establish a worldwide rate or mean that every connection caused an incident. DNSFilter sells network-security tools.
Visibility itself has limits. A November 9 research preprint examined how to detect unapproved agents on Windows endpoints. Its controlled evaluation used one workstation, 12 synthetic agents and nine benign workloads. The authors warn that the small test cannot rank detection products. It does, however, illustrate why a process or network connection alone may not tell a defender whether software is an AI agent or what it did.
A domain list may show some traffic from managed devices, while missing activity on a personal phone or AI features inside an approved service. An inventory of connected apps may reveal permissions but not everything a person pasted into a separate account. No single log is a complete map, and monitoring should respect applicable law, privacy commitments and workplace policies.
What employees can do now
- Check the rule and the account. Find out which tools, accounts and uses are approved. A familiar brand or a paid subscription does not by itself make a personal workspace company-approved.
- Keep sensitive material out until the use is authorized. Do not paste customer records, private documents, source code, passwords or API keys into a personal tool unless your organization has approved that specific use. Ask for a safe alternative if the task still needs to get done.
- Read connection requests. Before allowing access to work email, files or calendars, check what is requested and use your organization’s approval process. Do not connect an agent or run a suggested command just because an AI tool recommends it.
- If you already shared sensitive information, report it promptly. Tell the security or privacy contact which service and account were involved, what kind of data was shared and when. Preserve relevant details; do not forward the sensitive content again to explain the problem.
- If a credential was exposed, act on the credential. Notify the owner or security team and revoke or rotate it through the issuing service as appropriate. Follow the organization’s incident process rather than assuming deleting a chat removes all copies or access.
What employers can do this week
- Ask what work needs help. Invite employees to name tools and use cases without demanding they paste confidential prompts into a survey. Provide a clear reporting route and avoid making disclosure harder through unnecessary blame.
- Offer an approved route that works. Give staff practical options for common tasks, define which types of data are allowed and make requests for new capabilities timely. Approval and a paid license alone do not eliminate security risk.
- Inventory connections and owners. Review authorized and unexpected applications, their permissions, owners, connected data and credentials. Identify whether an agent can read, write, send or execute, and require human review for consequential actions.
- Limit and review access. Apply least privilege and approval controls to integrations; remove access that is no longer needed. Microsoft’s app-consent guidance explains how delegated permissions and organization-wide grants differ. Verify the actual grant and follow the provider’s incident procedure before changing access.
- Prepare for exposure. Know who can investigate, preserve relevant logs, revoke grants and credentials, contact a vendor about retention or deletion, and assess notification duties. Stopping an active compromise takes priority; do not delay containment while building a perfect inventory.
Our guides to protecting a business from hackers and responding to exposed API keys can help turn those steps into a workable plan.
The uncomfortable question for a company is not simply whether it has an AI policy. It is whether employees can do useful work within that policy, and whether the company knows where its information and permissions have gone. An approved tool can still be configured badly. A ban can still be ignored. Neither is a substitute for understanding the work, the data and the access involved.