For every human who logs into your cloud environment, roughly 100 non-human identities are already there – service principals, API keys, OAuth tokens, and AI agents. Your IAM program was designed for the human login. It was not designed for the service account with admin rights to seven production databases that hasn’t been rotated in fourteen months. The problem is not that you have too many identities to manage. It is that 99 of every 100 are governed by controls built for the one that is human.
The Category Error in Your IAM Program
The Cloud Security Alliance’s 2026 State of Cloud and AI Security documents what most security teams already sense but haven’t confronted directly: you are no longer primarily defending human logins. You are governing a non-human identity perimeter where service principals, secrets, and autonomous agents outnumber human users by a ratio of 100-to-1. Palo Alto Networks’ 2026 Identity Security Landscape puts the average at 109 machine identities per human, and both agree on the direction up and fast.
Every control in a traditional IAM program was built around properties that only humans have. Multi-factor authentication assumes someone can receive a prompt. Password policy assumes a password that gets typed. Phishing training assumes a person who can be deceived. Quarterly access reviews assume a manager who recognizes their reports. Offboarding assumes an employee who leaves. Not one of these assumptions applies to a service account or an AI agent.
The 100-to-1 ratio is not a scale problem. It is a category error. The overwhelming majority of identities in your environment are being governed by a model designed for the small minority that happen to be human.
Propelex Security Intelligence
This is why the CSA now ranks the exposure of insecure identities and machine permissions as the single top cloud security risk of 2026. Attackers have noticed the asymmetry: they target service accounts and AI agents to move laterally through cloud environments, because those identities carry high privilege and almost none of the governance rigor applied to human accounts.
Why Machine Identities Break Every Human Assumption
Non-human identities do not merely need different controls. They actively break the ones built for humans.
They never leave. A human is offboarded when they resign. A service account is created for a pilot project, a proof of concept, or a one-time integration and then the project ends, the credential persists in a config file or environment variable, and no offboarding process ever fires. The CSA calls this ephemeral permanence: the agent is gone, but the credential is not. Only 20% of organizations have a formal process to offboard and revoke API keys at all.
They can’t be re-authenticated. When a human credential looks risky, you can force a step-up: another factor, a re-login, a verification prompt. A machine identity authenticating machine-to-machine has no one to prompt. The credential either works or it doesn’t, and if an attacker holds it, it works for them exactly as it works for the legitimate workload.
Their blast radius outlives their purpose. The CSA’s 2026 token sprawl analysis coined “persistent blast radius” for credentials that outlive their operational context and remain exploitable indefinitely. Combine that with privilege creep, keys that started narrow and ballooned through loose role definitions and a single orphaned service account can quietly hold total control long after anyone remembers creating it.
The AI Agent Is a New Kind of Insider
If the static service account broke the human IAM model, the AI agent shatters it. Agents are not passive credential holders running fixed scripts. They acquire permissions dynamically at runtime, spawn sub-agents, invoke external APIs, write and execute code, and chain actions across dozens of systems in a single task.
These agents are the new insider threat. If an agent is overprivileged, an attacker can use it to exfiltrate data at machine speed without ever compromising a human credential.
CSA 2026 State of Cloud and AI Security
The scale of this is not hypothetical. Organizations expect AI agent identities to grow 85% over the next twelve months. And here is the governance gap in a sentence: most security leaders can explain what their AI agents are for, but far fewer can define what those agents can actually access, how that access is limited, when the permissions are revoked, or which systems can inherit them. The agents already reach financial records, PII, operational technology, and core business systems. The controls to bound them mostly do not exist yet.
There is a revealing split in the data. C-suite executives tend to believe their organizations enforce least privilege well because they are looking at human access. Security practitioners disagree, because they see the machine layer, where broad standing permissions are routinely left in place to keep operations running. Only 15% of organizations feel highly confident in their ability to prevent NHI-based attacks. The gap between the boardroom’s confidence and the practitioner’s reality is exactly where the 2026 breaches are happening.
Why “Discover Everything” Already Failed
Most non-human identity programs begin with a reasonable-sounding goal: discover all the identities first, then govern them. In practice that goal is already obsolete, and teams often realize it only after investing heavily in an inventory-first approach.
The problem is structural. Ephemeral Kubernetes service accounts, short-lived CI tokens, GitHub runner identities, and AI agent credentials do not sit still long enough to be inventoried. They are created through code, configuration files, CI/CD pipelines, and agent frameworks, channels designed for velocity, not for the governance metadata an identity system needs. A quarterly audit is a snapshot of yesterday’s problem.
The 2026 Trivy and GitHub Actions supply-chain compromises made the lesson concrete. Attackers did not exploit a dormant, discoverable account. They stole runtime credentials from CI runners, reused them within minutes, and fanned out across dependent pipelines before defenders could react. None of it was caused by a weak password or an inattentive user. It was the direct result of unbounded machine trust and no inventory, however complete, would have caught a credential that lived for minutes.
The Five Controls That Actually Work
Governing the non-human perimeter means abandoning the human-centric playbook, not scaling it up. Five controls address the crisis directly.
- Replace static secrets with ephemeral, identity-based credentials. Long-lived API keys are the root of persistent blast radius. Authenticate workloads through a verified non-human identity framework so that even if a component is compromised, the attacker’s window is minutes or seconds, not months.
- Enforce just-in-time access for agents and workloads. Standing permissions on an AI agent are a gift to whoever compromises it. A credential should be tied to a specific authorized task and expire the moment that task finishes. If it isn’t scoped and time-bound, it shouldn’t exist.
- Assign every non-human identity an accountable owner. The orphaned account problem is a governance problem. No credential should exist without a named human or team responsible for its lifecycle, its scope, and its retirement. Ownership is what makes offboarding possible at all.
- Automate rotation and revocation. With 24% of organizations taking more than a day to revoke an exposed credential, manual processes are the failure. Rotation that doesn’t break production and revocation that fires on detection, not on a human’s schedule are the baseline for machine-speed identities.
- Deploy identity threat detection built for machines. Behavioral analytics that flag when a machine identity acts out of pattern, reaching systems it never touches, at volumes it never generates is the only monitoring that works when there is no login to inspect. Self-narrating agent behavior is a detection opportunity; use it.
The Bigger Picture
The security perimeter has quietly relocated. For decades it ran around human users, authenticate the person, and you controlled access. In 2026 the perimeter is the sprawling web of non-human identities that actually run the business: microservices talking to databases, functions triggering automations, agents making decisions at machine speed. The OWASP Top 10 Non-Human Identity Risks, published in 2025, was the first mainstream framework to say plainly what practitioners already knew, machine identities are now the primary attack surface.
The organizations that adapt are the ones that stop treating non-human identity as an extension of the human IAM program and start treating it as its own discipline, with its own lifecycle, its own assurance levels, and its own controls. The ones that don’t will keep running MFA and quarterly reviews against the 1% while the 99% sit unmanaged.
Your most powerful identities do not log in. They do not forget passwords. They do not fall for phishing. And they now sit at the center of the most damaging breaches of the year. Governing them requires admitting the old model was built for someone else.
Propelex Security Intelligence
The machine identities already outnumber your people 100 to 1, and that ratio is only widening. The question is not whether to bring them under governance. It is whether you can enumerate, scope, and contain them before an attacker finds the orphaned service account you forgot you created.
Propelex helps organizations bring the non-human identity perimeter under control, inventorying service accounts, API keys, and AI agents, mapping their real privilege and blast radius, and rebuilding around ephemeral credentials, just-in-time access, and machine-aware detection. We start with a Cybersecurity & Privacy Risk Assessment scoped to your identity fabric and pair it with Privileged Access Management for the credentials that matter most.
