Immuta CISO warns of critical AI agent incidents at major enterprises

Michael Scott, Chief Information Security Officer at Immuta, predicts that at least two major enterprises outside the artificial intelligence sector will suffer critical AI-related incidents within the next 18 months. According to an interview with Help Net Security, Scott attributes this risk to organizations failing to properly manage permissions and identities for autonomous AI agents.

Scott argues that AI agents should be treated as lower-trust identities rather than inheriting the full access rights of the human users who deploy them. He warns that agents often leverage access to resources not intentionally targeted by users, such as available APIs, leading to security gaps that traditional controls miss.

The Mechanics of Agent Privilege Creep

Privilege creep for AI agents differs significantly from the gradual accumulation of access seen with human employees. While human privilege creep typically results from role changes or temporary project needs that persist, agents introduce a level of unpredictability. An agent might operate on behalf of multiple employees with varying permissions, complicating access decisions.

Agents can also create sub-agents to handle specific parts of a task, further expanding their operational scope. This behavior means access decisions are often made dynamically, without the oversight present when the original agent was deployed. Scott notes that agents may find ways to accomplish tasks that security teams did not anticipate when permissions were originally granted.

Identity and Authentication Gaps

The primary concern for Scott is determining whether an agent operates under a dedicated identity or relies on the user’s permissions. He states that it is not uncommon for agents to have excessive permissions, particularly when they use user credentials that vary across different staff members. Authentication alone does not secure what an agent does once it is running.

Scott emphasizes the need for strict control over agent workspaces. He advocates for permissions that disappear once the agent’s work is finished, rather than retaining broad access for potential future use. This approach requires a shift in how organizations view machine identities compared to human ones.

What should you do?

Immuta addressed these risks internally by investing in a multi-model AI platform with managed guardrails. This workspace supports employee-built agents while reducing the risk of excessive privileges. The company’s approach focuses on internally managed controls to ensure agents do not retain access beyond what is necessary for their immediate tasks.

What to do and how to stay safe: Immuta

• Audit AI agent identities to ensure they operate under dedicated credentials rather than inheriting user permissions.
• Implement strict least-privilege models where agent access is revoked immediately after task completion.
• Monitor agent behavior for unexpected API calls or resource access that deviates from intended workflows.
• Deploy internal guardrails and workspace controls to limit the operational scope of employee-built agents.

Step-by-step guide: Prompt Injection Warning Signs: How to Spot AI Manipulation

General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

Why are AI agents considered higher risk than human users for privilege creep?

Agents introduce unpredictability by potentially creating sub-agents and leveraging access to resources like APIs that humans do not intentionally target.

How long does the Immuta CISO predict it will take for critical AI incidents to occur?

Michael Scott predicts at least two critical incidents at major enterprises outside the AI sector will occur within 18 months.

What is the recommended approach to handling AI agent permissions?

Organizations should treat agents as lower-trust identities with dedicated permissions that do not inherit everything the user can access.