Salta al contenuto principale

AI agents need boundaries, not blind trust

Students collaborating on laptops, illustrating AI agent security risks, cybersecurity awareness, and digital learning.

Condivi questa pagina

Cecelia Taylor
Cecelia Taylor

Your team is probably deploying AI agents right now. The harder question is what those agents can reach once they're running. In episode 82 of The STEMINISTS Podcast, "Your AI Agent Just Became a Security Risk. Now what?", hosts Phoebe Goh and Darnell Fatigati chat with Sandra Dunbar, the NetApp® lead for cyber resilience product marketing. Sandra's answer is practical, and it stuck with me: the future of AI security isn't about trusting agents more. It's about trusting them less, intelligently.

Listen to episode 82 of The STEMINISTS Podcast

The STEMINISTS podcast featuring Phoebe Goh discussing AI agent security risks, cybersecurity, and responsible AI innovation.

Key takeaways

  • Excessive access makes well-intentioned agents risky
  • Machine-speed errors demand guardrails and emergency stops
  • MCP is a critical trust boundary
  • Zero trust requires identity, limited access, and monitoring
  • Immutable data enables recovery when controls fail

What is the biggest security risk with AI agents?

  • The biggest risk is excessive access, not malicious intent. As Sandra Dunbar puts it, "The greatest risk usually isn't a malicious agent. It's a well-intentioned agent that has too much access."
  • An agent that holds broad permissions and unclear guardrails can do real damage while it tries to help. Sandra points to the PocketOS incident, in which an agent without proper trust boundaries deleted a production database and its backup. As she describes it, the agent wasn't trying to destroy anything. "[I]t was trying to complete the task, but it had more authority than it had judgment."

Why do AI agents make mistakes at machine speed?

A person might make one bad call. An agent can make the same one thousands of times—in seconds. Sandra explores what that speed changes, why we should assume compromise, and where an emergency stop fits in.

How does the Model Context Protocol change the risk?

The Model Context Protocol (MCP) standardizes how agents connect to tools and data. That convenience creates a high-value trust boundary, and you need to secure it on purpose. In Sandra Dunbar's words, "MCP gives agents a doorway into your tools and your data, and the security challenge is making sure that that doorway doesn't become a shortcut to your existing control."

Does zero trust apply to AI agents?

Yes, and Sandra argues it applies more strictly than it does for people. Every agent needs an identity, permissions, monitoring, and accountability, the same as any employee or application. She describes a layered approach:

  • Identity verification: Confirm which agent is acting before it touches anything.
  • Role-based and attribute-based access control (RBAC and ABAC): Limit each agent to the data and actions its task requires.
  • Data-layer access controls: Restrict access where the data lives, not only at the application.
  • Behavioral monitoring: Extend user behavior analytics, originally built for human users, to cover agent activity.
  • Human approval for high-risk actions: Agents can't receive multifactor authentication prompts directly, so route sensitive actions through approval workflows and administrator alerts.

What happens when controls fail?

Even strong controls have limits. When an agent crosses a boundary, immutable, isolated data copies can provide a route back. But what does effective recovery look like when AI mistakes happen at machine speed?

What do teenagers have to do with AI agents?

Perhaps more than you'd think. Sandra draws a surprisingly apt connection between raising an independent teenager and managing AI agents: freedom works best when it expands gradually, within clear boundaries.

Her analogy gets even more interesting when friends—or additional agents—enter the picture. Listen to the episode to hear why managing a group of overconfident teenagers might be useful preparation for securing multiagent systems.

Hear the full conversation

I encourage you to listen to the full conversation. Phoebe, Darnell, and Sandra cover all of this in under 15 minutes. Listen to episode 82 of The STEMINISTS Podcast and follow the show so you catch every new episode. Leave a review or connect with Darnell Fatigati and Phoebe Goh on LinkedIn.

Learn how to secure data, reduce risk, and recover quickly from attacks with an infrastructure architected for resilience.

FAQ

What is the biggest security risk with AI agents?

The biggest security risk with AI agents is excessive access, not malicious intent. In this STEMINISTS episode, NetApp’s Sandra Dunbar explains that a well-intentioned agent with broad permissions and weak guardrails can cause significant damage because it acts at machine speed without human judgment.AI agents act at machine speed without human judgement, so a single bad decision can repeat thousands of times in seconds.

Does zero trust apply to AI agents?

Yes, zero trust applies to AI agents. Each AI agents need an identity, permissions, monitoring, and accountability, with identity verification and monitoring should be even stricter than for human users.

How should enterprises apply zero trust to AI agents?
Enterprises should apply zero trust to AI agents through layered controls that limit what each agent can reach and watch what it does. Sandra Dunbar of NetApp outlines five layers:

  • Identity verification: confirm which AI agent is acting before it touches any system or data.
  • RBAC and ABAC: use role-based and attribute-based access control to restrict each AI agent to the data and actions its task requires.
  • Data-layer access controls: enforce restrictions where the data lives, not only at the application.
  • Behavioral monitoring: extend user behavior analytics, originally built for people, to track AI agent activity.
  • Human approval for high-risk actions: AI agents can't respond to multifactor authentication prompts, so route sensitive actions through approval workflows and administrator alerts.

What is the Model Context Protocol (MCP) and why does it matter for security?

The Model Context Protocol (MCP) is a standard that defines how AI agents connect to tools and data. That makes it a high-value trust boundary, so organizations need to secure it explicitly. Without identity verification and least-privilege access at the MCP layer, an AI agent can bypass the controls you already have in place.

How do you recover from an AI agent mistake?

You recover from an AI agent mistake by restoring from immutable, isolated copies of your data. They give you a recovery path when an agent deletes or changes something it shouldn't. Zero trust controls reduce risk, but they can still fail. When an AI agent deletes or changes data it shouldn't, immutable, isolated data copies give you a clean path back.

Cecelia Taylor

Cecelia Taylor

A lifelong Northern California resident, Cecelia has been a marketing communication professional for 20+ years. She previously held social media roles at several Silicon Valley organizations including 8x8, Mellanox Technologies and Cisco managing social media strategy and metrics. Prior to her career in social media, she worked in audience development marketing for B2B publishing. She holds a BA from Mills College in Oakland, CA. She has two adult children and likes to read and loves to rock holiday sweaters during the season.Visualizza tutti i post di Cecelia Taylor

Passi successivi