Shadow AI agents are already working inside most large companies, and most security teams cannot name them. About 1,280 commercial business products now embed autonomous AI features, and only 282 of them sit behind single sign-on, according to telemetry reported in the 2026 State of Agent Security research and summarised by SparTech Software. The other 1,000 or so operate outside the identity system the company spent a decade building.
That gap is the story. The agents were not bought in a separate procurement round. They arrived as updates inside software the company already used, for customer support, sales, project tracking and code review. A separate analysis of the same research, published on 11 October, describes them as unmonitored identities holding long-lived tokens and broad access to mailboxes, files and databases. They can read. In many cases they can act. They do not appear in the audit log under a person's name.
The numbers behind shadow AI agents
The scale is not marginal. Telemetry analysed by Reco found that four in five AI tools operate without formal IT oversight, at an average of 414 unsanctioned AI tools per 1,000 employees in some environments, as reported by SparTech Software. Of 500 published agent tools examined in that research, 62 percent could both read local data and reach the internet, and about half could execute shell commands.
Worker behaviour explains part of the growth. An analysis drawing on the Okta AI Agents at Work 2026 survey reported that 67 percent of US workers use unapproved AI tools for their jobs, that the median time to detect an unauthorised tool is 403 days, and that only 11 percent of AI applications are visible to IT even though 90 percent of executives express confidence in their visibility. The gap between confidence and visibility is the operational risk.
Vendors are now selling the opposite model. On 8 October Google announced a persistent Gemini agent for enterprise work that can run for hours, keep context across systems and delegate subtasks, with dedicated identities, permissions, auditing and spending controls, as summarised in a weekly roundup of the announcement. Identity and audit are becoming product features for the agents a company chooses. They remain absent for the agents it did not choose.
Why shadow AI agents are a knowledge problem
Security teams frame this as an access problem. It is also a record problem.
When an agent summarises a ticket, drafts a reply from a mailbox or updates a record in a customer system, it creates company knowledge. If that work happens under a shared service account, inside a vendor cloud, with no link to the person who delegated it, the company cannot reconstruct the decision later. It cannot show which sources the agent read, which version of a policy it used, or whether a person reviewed the output before it left the building.
That absence compounds. A record depends on provenance: the statement, the source behind it, the date, and the person accountable for it. Shadow work has none of those properties by default. The result is a second, invisible layer of company activity that search cannot index honestly and that a review process cannot correct, because no one knows it exists.
This connects directly to the risk examined in the blog post Memory Poisoning: The Attack That Turns AI Agent Memory Into a False Record. A false fact that enters the record can be found and corrected at the source. A false fact handled by an unnamed agent, in a tool outside the inventory, has no source to correct.
What a governed agent record requires
The fix starts with inventory, not prohibition. The same Forkast analysis noted research from the Cloud Security Alliance and Unseen indicating that providing approved AI alternatives can cut unauthorised use by 89 percent. Workers use unapproved tools when the approved path is slower or absent. A ban without an alternative moves the work further out of view.
Three conditions separate a governed agent from a shadow one. First, the agent holds no access apart from the person who invoked it, so its reach ends where that person's reach ends. Second, every action it takes is logged against that person, so the work can be audited and reversed. Third, knowledge it drafts does not publish itself. It waits for a person to review it, in the open, before it joins the shared record. Those conditions are the basis of the approach set out in How Jilu Works.
None of this requires stopping agent adoption. The agent count will keep rising. The question for a company is whether that work leaves a record it can read, check and correct, or a shadow layer it discovers only after an incident, an audit or a departure forces the question. A company brain cannot be built from work no one can name. Inventory, identity and review are where it starts.