Your Employees’ Personal AI Agents Are Becoming Your Security Problem
David Goldschlag is co-founder and CEO of Aembit, an identity control plane for AI agents.
gettyAhead of an important client meeting, a sales representative connects their personal AI agent to company email and files. The setup takes minutes and allows the employee to arrive prepared with considerably less effort, but that well-meaning shortcut may carry risks the rep never considered.
Personal AI agents, including Meta’s Muse, OpenAI’s Dots and Google’s Gemini Spark, are software assistants that can carry out routine tasks across connected services and continue working between conversations. It is early, and there is little data yet on how often employees connect these personal assistants to company accounts. But if the first numbers are any indication, the habit is forming quickly. Meta said Muse passed 10 million accounts within roughly six weeks of launch, and a Menlo Ventures survey found that 32% of U.S. adults who use AI have already let it act on their behalf without final sign-off.
For security leaders who remember the first requests to connect personal smartphones to company email, the arrival of these agentic deputies has a familiar feel. Employees found something useful and expected to use it, while IT and security were left to reconcile that enthusiasm with responsibility for company data.
Bring-your-own-device (BYOD) policies and years of “shadow IT” disputes established an enduring reality: The rank and file would have a say in the technology they used. But the analogy has limits in the AI era. Traditional software generally follows predefined workflows, while an agent can choose which tools to use and which steps to take in carrying out an employee’s request, without the employee directing each action.
Permissions granted with little consideration during setup can later support a series of actions the employee never specifically contemplated. For CISOs, the difficulty is deciding how much authority should accompany that discretion and how to enforce those limits once an AI assistant begins working.
Consider the employee preparing for the customer meeting. The request of their agent sounds innocuous: Review recent correspondence, find the relevant account documents and prepare a briefing.
Yet the permissions available during setup may cover far more information than this assignment requires. Depending on the connection, the agent may receive access to a mailbox containing correspondence with hundreds of customers, along with internal discussions unrelated to the meeting.
Suppose the employee also asks the agent to prepare and send an external follow-up. An internal document contains useful background, including an unannounced pricing change. The agent could incorporate that information into a customer message because it appears relevant to the assignment. Every connection involved might be permitted, even though the resulting disclosure was never intended. Read-only access offers limited reassurance if the agent can send what it reads through another service.
Requiring employee approval before an agent receives access provides another safeguard, but its value depends on what the employee can reasonably assess. A consent screen can describe the permissions being requested without making clear how an agent could use them together. The employee may approve those permissions without anticipating the disclosure described above.
Company policy must establish which resources an assistant may access and which actions require further review. Employees should understand their responsibilities, but connecting an agent should not leave them to determine how much authority the company is willing to delegate.
The following practices can help security leaders establish those boundaries while giving employees a practical way to use personal AI agents.
Verify that records identify which agent acted and whose authority it used. Logs that name only the employee may leave investigators unable to distinguish direct actions from those taken by an agent.
Establish which accounts and resources it can access, whether it can act in the background and how the provider handles credentials and retrieved data. Approval of an assistant should not imply approval of every connection it supports.
Where controls permit, restrict access to relevant resources and separate retrieval permissions from permission to publish or send. Enforce those limits at runtime, when the agent requests access. If a service offers only broad access, weigh that limitation in the approval decision.
Reserve employees’ attention for consequential decisions. Before an external message is sent, for example, let them review its recipients, attachments and exact content, with another review required if material details change.
Ensure security can revoke an agent’s access without disabling the employee’s account. This gives security teams a way to intervene when an agent needs to be stopped. Here, that means access revocation: withdrawing the agent’s ability to use connected systems. Cutting off access, however, does not necessarily terminate the agent’s process.
Test it against background activity and previously issued credentials, and measure how long it takes for subsequent requests to fail. Depending on the implementation, previously issued access tokens may remain usable until they expire, or a revocation decision may take time to reach every system involved. Incident procedures should account for the behavior observed in testing, including any additional steps required to contain ongoing activity.
These practices also expose limitations that a policy document alone cannot resolve. If a provider cannot support sufficiently narrow access or useful audit records, the company may need to restrict the information available to that agent or decline the connection.
The BYOD experience offers a useful lesson about AI adoption. Employees’ preferences can influence company technology decisions long before a formal policy catches up. Security leaders should take seriously the reasons people want personal agents, including the convenience of using a tool they already know.
A workable policy should give employees a clear route to request access, explain which uses are permitted and offer alternatives when a requested connection grants more authority than the business can accept. Security teams, in turn, need evidence that the agreed restrictions hold during actual use.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?