HomeMalware & ThreatsAI Agents Are Using Employee Identities

AI Agents Are Using Employee Identities

Published on

spot_img

Governing AI Agents: A Call for Enhanced Identity Security

On November 18, 2026, a significant dialogue emerged surrounding the governance of artificial intelligence (AI) agents, particularly concerning their identity security. Jyotsna Raghunathan, a principal product acceleration specialist at Okta, highlighted alarming statistics indicating that approximately 95% of enterprises lack comprehensive visibility into the AI agents operating within their environments. These AI agents often operate within employee sessions or service accounts without possessing their own distinct identities. Should any of these agents act inappropriately, the repercussions might necessitate locking access for the employee themselves, representing a substantial risk in securing AI agents.

The Zero Trust Paradigm

Raghunathan emphasized the evolving necessity of the Zero Trust security model, initially developed to verify every individual and device, now including AI agents as an essential component. This shift aims to mitigate the inherent risks associated with agents borrowing identities from human users. The result is a precarious situation where uncertainty breeds hesitance, confining many agents to pilot phases. The sentiment was echoed by Scott Sikorski, a principal at PwC, who noted that the reluctance to transition AI agents into full production stems from an atmosphere of fear, uncertainty, and doubt. To move agents beyond trial runs, an emphasis on identity security is imperative, requiring that every AI agent possesses its own identifiable credentials.

The Need for Distinct Identities

Imagining an agent working within a user’s browser, executing tasks such as filing expense reports, reveals the potential for confusion. To every system it interacts with, the agent masquerades as the employee, posing significant security concerns. Raghunathan argued that it is crucial for agents to have their own non-human identities registered within the directory of record. The responsibility of overseeing the onboarding and management of these identities should fall to whoever controls that directory.

Amidst this backdrop, John MacDonald, Google’s global director for Workspace growth strategy and solutions, presented the orchestration platforms like Google’s Gemini Enterprise, which serve as frameworks for governing and monitoring agents. He emphasized that while these platforms provide the necessary structure, foundational identity management—such as that provided by Okta—is essential for anchoring these agents.

Setting Boundaries for AI Agents

The expansive access granted to agents poses another layer of complexity. MacDonald analogized the dual capabilities of an agent to that of a chief of staff, assisting executives by aggregating critical information across emails, chats, documents, and meetings—a breadth of insight that could otherwise be lost. However, the necessity for strict access limitations becomes evident; AI agents should have access strictly within the bounds of the least privilege principle. This principle dictates that agents should not have more access than the users they represent, and, in most cases, their permissions should be more restricted.

Visualizing permissions in a Venn diagram illustrates this concept: the overlapping areas between user permissions, agent permissions, and corporate policies must dictate where an agent can operate. For instance, agents acting on behalf of sales personnel might require human approval to authorize purchases exceeding a certain threshold.

Implementing Accountability Measures

Beyond identity security, effective governance necessitates robust tracking and accountability mechanisms. Sikorski underscored the importance of maintaining an action record that documents who requested specific actions, which agent executed them, and the nature of those actions. This documentation is considered the minimum standard expected during audits. Managed by the Cross App Access standard championed by Okta, these records can link back to both the user and the agent in question.

The creation of a tangible means to disengage an agent—that is, a "kill switch"—is just as critical. Raghunathan argued that without a dedicated identity for each agent, revoking access becomes a complicated process. Moreover, monitoring agent behavior is vital; unexpected actions from agents can raise red flags, allowing organizations to revise policies or suspend activities proactively.

Assigning Responsibility and Governance

Ultimately, accountability for AI agents should reside with the teams responsible for their intended business outcomes. MacDonald insisted that these teams must initiate changes if an agent fails to meet expectations. Sikorski further posited that effective governance of agentic AI will require collaborative efforts across identity and access management frameworks in tandem with AI governance and business teams to determine which agents should be deployed.

Certification processes must also run both ways: ensuring that the resources an agent connects to remain relevant and that users still require the agent itself. Raghunathan noted that only the technical and business owners directly involved can provide these crucial answers.

A Strategic Approach for Chief Information Security Officers (CISOs)

As organizations prepare to deploy new AI agent security measures, CISOs are urged to focus on foundational governance elements. Sikorski suggested initiating the next 90 days with essential actions, including defining governance structures, inventorying existing agents, and establishing policies before selecting enforcement tools. Engaging experienced partners to avoid the pitfalls of isolated leadership in AI governance—an “island of one”—is crucial for expediting the development of a comprehensive framework.

Selecting pertinent use cases to secure basic productivity scenarios is advised, followed by choosing scalable platforms as agent governance standards continue to evolve. MacDonald has found that successful projects often utilize an approach of launching at around 80% completed infrastructure, using user feedback for continuous improvement.

As organizations navigate this complex landscape, the dialogue surrounding AI agents, identity security, and governance will remain a priority, ultimately driving the next phase of technological advancement and organizational security.

Source link

Latest articles

JPMorgan Chase Confirms Cyberattack on Third-Party Vendors with Potential Exposure of Customer Data

Cyberattack Targeting Third-Party Vendors Raises Concerns for JPMorgan Customers In a troubling revelation, JPMorgan Chase...

JPMorgan Confirms Data Breach Involving Customer Information Due to Third-Party Compromise

JPMorgan Chase has recently confirmed a significant data breach resulting from a cyber-attack targeting...

Live Webinar on Closing the Agent Trust Gap: Bridging the Divide Between Agent Ambition and Identity Readiness

Jenny Feng: A Leader in Cyber Risk and Compliance at PwC US Jenny Feng serves...

Is AI Dulling Skills or Creating Smarter Cybersecurity Professionals?

Is AI Dulling Skills — or Creating Smarter Cybersecurity Professionals? The rapidly evolving landscape of...

More like this

JPMorgan Chase Confirms Cyberattack on Third-Party Vendors with Potential Exposure of Customer Data

Cyberattack Targeting Third-Party Vendors Raises Concerns for JPMorgan Customers In a troubling revelation, JPMorgan Chase...

JPMorgan Confirms Data Breach Involving Customer Information Due to Third-Party Compromise

JPMorgan Chase has recently confirmed a significant data breach resulting from a cyber-attack targeting...

Live Webinar on Closing the Agent Trust Gap: Bridging the Divide Between Agent Ambition and Identity Readiness

Jenny Feng: A Leader in Cyber Risk and Compliance at PwC US Jenny Feng serves...