Security teams at large companies face a question few can answer with confidence. If an autonomous AI system pulls data from core repositories or modifies production code today, who originally approved its actions? The employee who built it might have left months ago. Yet the agent keeps running. Its permissions stay intact.
This scenario plays out more often than leaders admit. The Hacker News detailed the problem in stark terms last week. Orphaned agents, those AI tools left active after their creators depart, retain standing privileges. They hold permanent, unrestricted access to sensitive databases and source code. Traditional security tools miss them because they treat these systems like ordinary applications. They cannot trace the borrowed identity back to a departed owner.
But the issue runs deeper. Recent industry reports paint a troubling picture. According to analysis from Iternal, 79% of organizations lack visibility into their AI agents and related systems. Another 47% see enterprise AI activity routed through personal accounts outside standard identity controls. And roughly 97% of non-human identities carry excessive privileges. Just 0.01% of those identities control 80% of cloud resources. The numbers come from surveys and audits compiled in the organization’s 2026 security checklist.
Token Security describes the fallout clearly. “As AI agents proliferate across enterprises, many become orphans, agents that remain active without clear ownership or oversight,” the company states on its site. “These agents remain active with access rights, but with no one accountable for their use. They pose serious risks, including hidden backdoors, static credentials, and compliance violations.” Attackers can exploit stale credentials without detection. Operational errors go untraced. Audits fail on over-privileged accounts.
Orchid Security’s work on the 2026 Identity Gap Report adds context. Only 57% of enterprise applications remain visible to a central identity provider. Some 67% of non-human identities authenticate locally, outside any identity provider. Forty percent of accounts sit orphaned. The report, discussed in a recent Techstrong TV interview with Orchid CEO Roy Katmor, highlights how agentic AI amplifies “identity dark matter” – unseen accounts, flows and permissions now callable by autonomous systems at scale.
Enterprises rushed these tools into production. Developers built agents to automate code reviews, data analysis or workflow orchestration. Then those developers moved on. The agents did not. Permissions lingered. Tokens never expired. And security teams lacked the mapping to connect actions back to responsible humans.
So what does an orphaned agent actually look like in practice? It might be a custom script using large language models to query internal wikis and generate reports. Or an autonomous system hooked into version control that can push changes without fresh approval. When the creator’s access gets revoked, the agent’s token often survives. It keeps operating under inherited rights. An attacker who compromises the agent gains persistent, low-and-slow entry.
The problem compounds with shadow AI. Teams spin up agents without central review. A 2026 Gravitee survey found only 24.4% of organizations maintain full visibility into agent communications. More than half run without security oversight or logging. The average company now manages 37 deployed agents, a number that grows quarterly. Shadow incidents cost an extra $670,000 on average, according to IBM research cited in multiple analyses.
But recent coverage shows the conversation is shifting. A LinkedIn post from cybersecurity analysts this month tied orphaned agents directly to broader identity risks. Vendors like Strata and Maverics push just-in-time provisioning for agents. Identities get created on demand with time-to-live attributes, purpose limits and risk context. Once the task ends, the identity retires. No lingering credentials. No sprawl.
Token Security outlines a practical path forward. Its platform focuses on ownership discovery – inventory every agent and flag those without owners. Lifecycle governance follows. Assign responsibility, rotate credentials, retire inactive ones. Credential hygiene demands regular rotation and expiration. Accountability controls ensure each identity carries a defined purpose, owner and lifecycle aligned with identity governance practices. The company stresses contextual awareness across agents, secrets and services. Remediation happens at scale without guesswork.
Iternal’s checklist takes a stricter stance. Treat a managed agent identity as the core unit of inventory. Flag any agent missing a registry entry, owner or managed identity as critical. The guidance draws from CISA’s 2026 Five Eyes recommendations, OWASP’s Agentic Top 10 and NIST frameworks. It calls for distinct, short-lived, scoped credentials. Default-deny policies. Human-in-the-loop checks for high-impact actions. Tamper-evident logging. Continuous red teaming. Air-gapped containment where possible.
These steps matter because AI agents differ from traditional service accounts. They reason, delegate and act across domains. They improvise. Static permissions become dangerous when an agent can chain tools or call APIs autonomously. One compromised agent can cascade across systems. The blast radius grows fast.
Compliance teams already feel the pressure. Over-privileged access triggers audit failures. Untraced actions violate data handling rules. Regulators expect accountability. Yet many organizations still cannot list every agent touching sensitive information or name its owner.
The good news? Tools and processes exist to close the gap. Continuous discovery maps applications, authentication paths and unmanaged accounts. Runtime policy engines evaluate actions in context. Just-in-time identities limit exposure. Centralized registries track ownership. None of this requires ripping out existing infrastructure. But it does demand attention now, before the administrative debt grows larger.
Security leaders who treat AI agents as just another non-human identity miss the point. These systems operate dynamically. They inherit permissions aggressively. They act without constant oversight. Ignoring the orphans invites trouble. Mapping them, owning them and governing their privileges turns potential liabilities into controlled assets.
Recent X discussions echo the urgency. Cybersecurity practitioners noted the new form of technical debt on June 18, linking back to the original reporting. Others called orphaned agents the evolution of shadow IT. The consensus? Visibility comes first. Then ownership. Then strict lifecycle controls. Delay those steps and the hidden access risks only multiply.
Companies cannot afford to treat this as tomorrow’s problem. Agent adoption accelerates. The identity gap widens. But with focused discovery, clear accountability and dynamic controls, security teams can catch up. The agents will keep working. The question is whether organizations will finally know exactly who stands behind them.
WebProNews is an iEntry Publication