Most AI agents in the enterprise never touch single sign-on

Only a small fraction of the AI-enabled third-party software running inside enterprises is visible to the identity systems meant to govern it, according to analysis published by The Hacker News on 10 October 2026, drawing on the 2026 State of Agent Security Report.

In the environments studied for that report, roughly 1,280 third-party products now embed AI in some form. About 282 of them sit behind single sign-on, the central login service that most organisations use to control and log access to corporate applications. That leaves roughly a thousand AI-enabled products outside the reach of identity infrastructure by default.

The article is explicit that this is not a story about concealment. Nobody hid the remaining products, it says. The gap exists because an identity stack can only govern what authenticates through it, and most agents never do. If a vendor bolts an AI feature onto a product that a team adopted directly, or if an agent calls an application through an interface that does not involve a corporate login, the identity layer has nothing to record and nothing to enforce against.

The practical consequence described is a mismatch between where security teams look and where AI actually runs. Controls built for the AI that an organisation deliberately chose and deployed, such as an approved assistant or a sanctioned model provider, do not extend to AI that arrived inside tools the organisation bought for other reasons. The analysis frames this as the third-party agent problem: the agents that matter most for risk are often the ones nobody decided to adopt as AI at all.

The figures quoted are averages from environments studied for the report rather than a universal count, and the piece does not state how the environments were selected or how an embedded AI feature was identified. Readers should treat the ratio, roughly one in five AI-enabled third-party products behind single sign-on, as the headline finding rather than the absolute numbers.

The analysis describes the visibility gap as the clearest signal of the wider problem it sets out to examine. It does not claim that any of the unmanaged products have been abused, and no incident is cited.

For security teams, the immediate question raised by the figures is an inventory one: knowing which purchased software has quietly gained AI features, and which of those features can act on company data without ever presenting a login to the identity stack. Until that list exists, the analysis suggests, coverage figures based on single sign-on will describe only the smaller share of the estate.