Shane O'Donnell, Centric Consulting's Vice President of Cybersecurity, has two decades of experience in audit and cyber risk management.
getty
By the end of 2026, 40% of enterprise applications will feature task-specific AI agents—up from less than 5% in 2025, according to Gartner.
For organizations racing to capture the productivity gains of agentic AI, the jump signals opportunity. For security teams, it signals a problem most organizations are not prepared to handle.
Security teams have spent decades getting good at one version of the access question: Who has access to what, do they still need it and what happens when they leave? Least privilege, access reviews and credential life cycle management became table stakes.
While they still work, these best practices were built for humans, who are no longer the only—or even the primary—actors in your organizational systems. In most enterprises, non-human identities (NHIs) already outnumber humans significantly, and the gap continues to widen.
NHIs have existed in enterprise environments for some time. What's new is that business teams, rather than security teams, are now deploying them by the dozens, often without any security review.
In fact, about one in seven AI agents currently running in production environments received full IT security sign-off before deployment, according to a recent Gravitee research.
Ungoverned agents are gaining access that was never formally reviewed, operating under shared credentials and remaining in systems long after the workflow that required them has changed or ended. Gartner flagged failure to address AI agent identity and governance as one of the top cybersecurity trends to watch in 2026, and the risk only grows as deployments scale.
The risk became apparent last August, when attackers compromised OAuth tokens tied to the Drift AI chatbot integration used by Salesloft and gained access to the Salesforce environments of more than 700 organizations.
The breach went undetected for days because the attacker's queries were indistinguishable from legitimate chatbot activity. Enterprises could see the chatbot had access, but they couldn’t see what it was doing with that access. To security teams, it was just a trusted non-human identity doing exactly what it looked like it was supposed to do.
That pattern will only become more common as agent deployment accelerates.
For CISOs, CTOs and security teams, reviewing your security posture means extending the governance you already know to a new class of identities:
Before any technical controls can be applied, you need to understand what you have. To do so, build an inventory of machine identities, from AI agents to service accounts, API keys and automation scripts in the environment. Document who owns it, what it can access and whether that access is still necessary.
Many organizations skip this, either because they’re moving fast or because they don’t have a formal catalog of their deployed agents at all. But you cannot govern what you have not found. And without an inventory, access reviews become impossible, and ungoverned credentials continue to climb.
The same principles that govern privileged human accounts apply directly to AI agents. The frameworks just need to be extended. These include:
• Just-In-Time Credentials: Grant access only for the duration of a specific task and revoke it immediately after, eliminating the standing privileges that make compromised agents so dangerous.
• Ephemeral Tokens: These should expire automatically, limiting the window of exposure if a credential is ever stolen.
• Automated Attestation Workflows: Ensure that agent permissions are validated continuously rather than set and forgotten.
• PAM integration: Extend existing privileged access management platforms to govern AI agent credentials with the same rigor applied to human privileged users.
The common thread is least privilege. Agents should have exactly the access they need to perform their function, for exactly as long as they need it.
Even well-governed agents can be compromised, misconfigured or manipulated. AI agents represent an expanding attack surface that operates at machine speed.
When an agent behaves unexpectedly, your team needs to know as it's happening. Behavioral analytics and logging need to establish a baseline for normal agent activity and flag deviations in real time. Critically, response workflows should not require a human to review a ticket before acting.
Technical controls only hold if someone is responsible for enforcing them. This means someone should:
• Embed access governance into development pipelines and make security a condition of deployment rather than a retrofit.
• Review access for non-human accounts on the same cadence as human user reviews, with audit trails.
• Treat every agent deployment as an identity decision, not just an engineering one.
The frameworks and disciplines for governing access exist. The work now is extending them to identities that don't have a manager, a badge or an offboarding date, before an ungoverned one shows up in your incident report.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。