Ghost Credentials Are the Next Cloud Budget Line Item

The number that should be on every platform lead's desk this quarter is 244. That's how many nonhuman identities a single developer can accumulate, according to research from security architect Aleksandr Krasnov, and it's the ratio that turns identity governance from a Q4 line item into a board-level conversation. For any team about to sign a multi-year contract with an IAM vendor or an AI orchestration platform, this is the number that changes the math on build-vs-buy.

The Numbers

The setup, as Dark Reading reported, is deceptively small. One AI-enabled workflow agent, idle for 30 days, suddenly starts firing off API calls at odd hours. That triggered an anomaly investigation, and Krasnov, a distinguished security architect at Ducker Tech Consulting, pulled the thread. What he found underneath was a mesh of ghost credentials: tokens, agents, and service accounts sitting outside the traditional trust boundary, still capable of lateral movement and privilege escalation.

Krasnov declined to name the victim, but the shape of the incident matters more than the identity of the customer. Consider the economics from an attacker's perspective. Two weeks to run a phishing campaign, or five minutes to hunt exposed tokens. Krasnov's own framing is that the rational attacker takes the five-minute path. That collapses the cost curve of intrusion in a way that most security budgets were not built to price.

Then layer the 244 number on top. A team of 50 engineers, on a naive multiplication, is theoretically responsible for over 12,000 nonhuman identities. Most CISOs I've spoken to are still budgeting identity governance as if their attack surface scales with headcount. It doesn't. It scales with the product of headcount and NHI density, and that density is a runaway function of how many AI agents, CI/CD pipelines, and SaaS integrations a team has adopted since 2024.

Krasnov's informal testing of his new open source tool, NHI Hound, targeted companies of up to 2,000 employees. His stated cleanup horizon for those organizations: six to nine months. That's not a sprint. That's a full fiscal half, dedicated to untangling trust relationships that were mostly created by defaults nobody reviewed. Above 2,000 employees, Krasnov says the graph becomes "super, super messy" and the tool's current output risks being too noisy to act on. So the useful window for a proactive cleanup, using the tooling that's about to ship, is narrow: mid-market teams with the discipline to spend two quarters on it, before they cross the 2,000-employee complexity cliff.

What's Actually New

Nonhuman identity risk is not a new category. Service accounts have been a compliance headache since the SOX era, and every cloud IAM vendor has been selling "least privilege" as a value proposition for a decade. What's genuinely different this cycle is three things stacked on top of each other.

First, the AI agent is now the credential holder. The trigger incident wasn't a leaked developer laptop or a misconfigured S3 bucket. It was a workflow agent that had been sitting quiet for 30 days and then woke up. That's an entirely new failure mode. Agents get provisioned by product teams shipping features, not by IAM teams reviewing access requests. The trust path is created as a byproduct of shipping, and nobody owns the offboarding.

Second, the pivot chain now leads directly into identity providers. Krasnov's Black Hat session, part of the online program at the end of Black Hat USA 2026, plans to walk through turning a single leaked key into full cloud compromise by abusing chained trust into providers like Okta. That's the piece that should scare every VP Engineering who assumed their SSO investment was a security boundary. If the IdP itself is reachable via a ghost NHI, then the entire zero-trust story collapses to whatever the weakest agent's token permits.

Third, tooling is finally catching up to inventory. NHI Hound ingests identity data from Okta, GitHub, and cloud IAM platforms, classifies identities by trust-severity levels, and flags what Krasnov calls "critical" conditions, where a low-trust identity can effectively act as a super-admin because of implicit trust relationships nobody explicitly defined. That last phrase is the one to underline. The risk isn't the permissions you granted. It's the permissions that emerged from combinations of grants across systems, which no single admin ever authorized as a unit. This maps cleanly onto MITRE ATT&CK techniques around valid accounts and cloud pivot, but the discovery problem has always outpaced the detection tooling.

What's Priced In for Security Teams

Most mature security teams already know NHI sprawl is a problem. The Gartner-style talking points have been circulating for eighteen months, and every vendor at RSA 2026 had "nonhuman identity" on their booth banner. What's priced in: the general anxiety, the awareness that agents proliferate faster than humans, and the sense that CIEM tooling from the major cloud vendors is incomplete.

What's not priced in, and this is where I'd push back on the market consensus, is the operational cost of remediation. Six to nine months of dedicated cleanup work on a sub-2,000-employee company is a real number. That's a headcount decision. For a Series B fintech running lean, that's either a new hire on the platform team or a deferred product roadmap. The CFO is going to feel this before the CISO does, because the CFO is the one who signs off on the platform engineer requisition.

Also not priced in: the vendor lock-in dimension. If NHI Hound (or any equivalent tool) becomes the source of truth for your identity graph, you've just added another dependency that spans Okta, GitHub, and your cloud IAM. That's fine while the tool is open source and self-hosted. It's a different conversation the moment someone acquires the project or offers a managed version. Teams evaluating this category should be modeling the three-year cost, not the pilot cost.

Contrarian View

The consensus reading of this story is that every organization should immediately run an NHI discovery exercise. I'd argue the opposite for a specific segment: enterprises above 2,000 employees should probably wait.

Krasnov himself says the current output becomes overwhelming at that scale. "Too much mess and too much noise" is his phrase, and it points to a real dynamic. Running a discovery tool that surfaces thousands of critical trust conditions, without the operational capacity to remediate them, creates legal and regulatory exposure you didn't have before. Once your GC knows you have a documented list of 4,000 critical identity risks, the calculus around breach disclosure and reasonable-care standards shifts. There's a legitimate argument for waiting until the tooling matures, the visualization improves, and remediation playbooks exist, before you generate the artifact that puts you on the hook.

This is not a popular view in security circles, where the ethos is always "you can't fix what you can't see." But governance without capacity to act is worse than ignorance in some regulatory contexts. That's a conversation for the security leader and the general counsel to have together, not for the platform team to decide unilaterally.

The Stakeholder Question

The Head of Platform at any Series B or C company running more than a handful of AI agents should be asking their CFO this week whether the next identity governance budget line is a tooling spend or a headcount spend. Those are different answers with different three-year cost curves. The open source path costs a platform engineer's time for six to nine months. The vendor path costs less upfront and more later, once the tool becomes load-bearing. Neither option is free, and pretending this is a project that fits inside existing IAM budget is the mistake that shows up in the postmortem.

Key Takeaways

• The 244-NHI-per-developer figure means identity attack surface scales with the product of headcount and agent density, not with headcount alone. Budget accordingly.
• Attackers now choose a five-minute token hunt over a two-week phishing campaign. Any threat model that still weights social engineering as the primary vector is out of date.
• The critical risk isn't explicit permissions, it's emergent super-admin capability from chained trust relationships that no single admin authorized as a unit.
• Six to nine months of dedicated cleanup is realistic for sub-2,000-employee organizations. Above that threshold, current tooling produces more noise than signal, and waiting may be defensible.
• Teams evaluating NHI discovery tools should now be asking who owns remediation, not just who owns discovery, and whether the output creates legal exposure they aren't ready to manage.