AI agents built into workplace tools can act with the data access of the employee who created them, making it harder to tell what a person did and what their agent did. A Nudge Security report cited by Help Net Security says the result is a growing blind spot across business systems.
Watch Desk analysis
What happened
Nudge Security’s report draws on anonymised, aggregated telemetry from its customers’ environments. Among organisations with at least one AI agent, the median had seven, and around 44% had more than ten. The agents appeared across 15 platforms, with more than half of these organisations using agents on two or more.
Some agents have their own bot or integration accounts. Others use the OAuth grant, API key or shared connection of the employee who set them up. That can give an agent access to the same company data as its creator, while logs still show the employee’s identity. The report says that access may remain even after the employee changes roles.
Why it matters
These agents can run inside project trackers, CRMs and other cloud services, beyond the reach of some endpoint monitoring tools. When activity is recorded under an employee’s identity, security teams may struggle to distinguish a person’s actions from an agent’s, or know which agents are operating at all.
Nudge Security recommends finding agents inside the platforms where they are built, checking whose access they use, and reviewing and revoking OAuth grants that are no longer needed. It also advises rotating any API key or access token pasted into an AI prompt. Those are the company’s recommendations, based on its telemetry and analysis.
Our read
The awkward shift is from asking whether staff are using AI to asking what each agent can reach, and whose name appears when it does. An agent does not need to be malicious to create a permissions headache; inheriting broad access and leaving a fuzzy audit trail is quite enough.
The report gives security teams practical checks, but its figures describe Nudge Security’s customer environments, not every organisation. Still, “we didn’t know the agent had that access” is a poor incident response plan.
What to watch
- Whether workplace platforms make it easier to identify agents and give them distinct identities.
- Whether organisations audit agents’ permissions when employees change roles or leave.
- Whether security logs can reliably separate an agent’s actions from its creator’s.
Discussion spark: Should workplace AI agents have their own identities and narrowly scoped permissions by default, even if that makes them harder to set up?
Sources and evidence
- Source update (19 November 2026, 05:00 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.