By Russell Spitler, Co-Founder and CEO of Nudge
For most of the last two decades, security and infrastructure teams have measured the attack surface by counting things: servers, VMs, containers, endpoints, open ports. That model assumed the company bought the technology and the company’s own people deployed it. Neither is true anymore. Employees now choose, connect, and authorize the tools that get their work done, and each of those decisions creates access that never shows up in an asset inventory. We call that place the Workforce Edge: where work happens and risk now grows. My predictions for 2027 all follow from it.
The inventory that matters will be relationships, not assets
A CMDB can tell you every virtual machine you run and still tell you nothing about who can read your customer data. That access lives in OAuth grants, API keys, browser extensions, and service accounts that employees create in a few clicks. A sales rep connects a note-taking app to their calendar and mailbox. An engineer wires a coding assistant into a repository. Each one is a standing credential held by a third party, and none of them were deployed by IT. In 2027, the teams that do well will stop asking what they own and start asking which identity can reach what, through which grant. Teams that keep counting assets will produce a very accurate report about a surface attackers have already moved past.
AI agents will become the largest unmanaged identity class
An agent acts with a human’s credentials but without a human’s judgment, at machine speed, around the clock. Most identity programs have no answer for that yet, because they were built to manage people and, at the margins, service accounts. Employees are building agents the same way they adopt everything else: a few clicks, on a platform IT never reviewed, with permissions copied from whoever created them. I expect the first serious agent incidents in 2027 to come not from attacks on models but from agents doing exactly what the wrong party told them to, using access nobody remembered granting. The right model is the one we already use for new hires: scoped access, a supervised period, trust that expands as it’s earned, and a named human who is accountable. Testing an agent like software and shipping it will fail for a plain reason. A test report can’t predict what the agent does in the case nobody thought of.
Friction will decide more than policy
Employees take the path of least resistance. If the sanctioned tool takes three months of procurement and the unsanctioned one takes a single click with a work Google account, the policy has already lost, and nobody has to be careless for that to happen. Nobody signs up for an AI notetaker to hurt their employer. Blocking makes it worse, because usage moves to personal accounts and devices where security has no visibility at all. In 2027, the organizations that pull ahead will treat unsanctioned adoption as a demand signal, discover it quickly, and shorten the approved path until it is faster than the workaround. They will save the hard no for the cases that deserve it, which are fewer than most programs assume.
Network controls will keep losing ground, for a mechanical reason
When an employee authorizes an AI service to read their drive or inbox, the data moves from one provider to another. It never crosses the corporate network, so proxies and gateways inspect traffic that isn’t there. Domain names have stopped meaning much too, since a single platform can host both a trusted app and a throwaway one. Network controls still have value. They just see less of the picture every year. The control point is the grant: what scope it carries and who approved it. Expect 2027 to be the year that review, expiry, and revocation of grants become routine program work instead of an annual cleanup.
Offboarding will be where programs visibly fail
Accounts and grants outlive the people who created them. When an employee leaves, HR closes the identity, and the integrations that person authorized keep working, reading mail and files on behalf of someone who no longer works there. When one of those vendors is breached, attackers don’t need to exploit anything on the customer’s side. They use grants the customer already issued. That is how a supply chain compromise reaches companies that never signed a contract with the vendor at fault. In 2027 this gap will surface in audits and customer security questionnaires, where “we don’t know which tools process our data” is not an acceptable answer, but for most organizations is the honest one.
The attack surface moved to where employees work, and the inventory should follow it. Start with discovery of what people have actually connected, rank it by the data each connection can reach, and revoke what nobody is using.
