What is one specific practice or control you recommend for securing AI agents or models in organizations, and why?

Most answers emphasize strong controls—like identity, permissions, monitoring, and guardrails—while others highlight human oversight, data protection, and peer review as key practices for securing AI agents.

13 answers

Answers

  • Casey Ellis

    Principal · Tall Poppy Group

    Data Classification

    Casey Ellis recommends data classification and labeling, saying companies will be forced into it to ensure only certain types of data are available to users and especially to their agents.

    Security Weekly - A CRA Resource · Watch on YouTube
  • Satya Nadella

    Chairman and CEO · Microsoft

    Identity and Activity Logging

    Satya Nadella says his team built Agent365 and Perception to give every agent an identity and log all its activity, including tool calls and model calls, so policy can be set and agent behavior monitored.

    The answer starts at 5:16.

    The Twenty Minute VC (20VC): Venture Capital | Startup Funding | The Pitch · Listen on Apple Podcasts
  • Harry Stebbings

    Founder · 20vc

    Agent Identity Logging

    Harry Stebbings does not speak in this passage; the answer about agent identity and logging is given by another speaker, so there is no line of his to report.

    20VC with Harry Stebbings · Watch on YouTube
  • Rekha Shenoy

    Chief Executive Officer · BackBox

    Human-vetted remediation

    Rekha Shenoy recommends human-vetted remediation tasks, where the AI creates a remediation task and an expert vets it before running it, as a way to build trust in agentic systems.

    The Everything Feed - All Packet Pushers Pods · The Everything Feed - All Packet Pushers Pods · Listen at 18:10
  • Gilad Friedman

    Head of Sales - Agentic AI Security · Radware

    Architecture and Governance

    Gilad Friedman says the practice to start with is architecture and governance, describing it as infrastructure architecture, feasibility and control.

    SecurityCafé · SecurityCafé · Listen at 21:06
  • Michael Jenkins

    C.T.O. · Threatlocker

    Peer review

    Michael Jenkins recommends a solid peer review session where all code, whether written by a developer or an AI developer, gets reviewed by a developer, saying you have to have all code reviewed no matter who writes it.

    Security Weekly - A CRA Resource · Watch on YouTube
  • Brennan Baybeck

    SVP & CISO - Customer Success Services · Oracle

    Prompt and Output Filtering

    Brennan Baybeck recommends prompt and output filtering as a data protection control, describing it as detecting and redacting secrets, PII, credentials and source code as data goes in and out of models.

    SiliconANGLE theCUBE · Watch on YouTube
  • Lou E.

    Federal CTO · ColorTokens Inc.

    Asset-level permission controls

    Lou E. recommends focusing on asset-level permission controls so AI agents cannot jump out of their sandbox, such as an HR agent reaching the finance server.

    Fed Gov Today · Watch on YouTube
  • Jensen Huang

    Founder and CEO · Nvidia

    Aggressive Monitoring

    Jensen Huang recommends "true aggressive monitoring of agent activity" with everything auditable down to every object an agent touches, framing long-running agents as new insider risks that could, for example, fake the books.

    Resilient Cyber · Resilient Cyber · Read
  • Denis O'Shea

    Founder & CEO · Mobile Mentor

    Identity and Least-Privilege

    Denis O'Shea says every agent should have an identity and be registered, with permission to access some databases and data categories but not others, following the permissions granted to its owner.

    Leadership Beyond Borders · Leadership Beyond Borders · Listen at 24:33
  • Maor Hod

    Co-Founder and CEO · Dash Security

    Runtime Protection Guardrails

    Maor Hod recommends runtime protection guardrails for each AI interaction, saying they let customers block threats while adopting AI with confidence without slowing the organization.

    SiliconANGLE theCUBE · Watch on YouTube
  • Rob Allen

    Chief Product Officer · Threatlocker

    Default Deny

    Rob Allen recommends a deny-by-default approach, saying it limits the blast radius by only allowing access to the people who need these tools, and that this applies to every organization.

    Business Security Weekly (Video) · Business Security Weekly (Video) · Listen at 9:47
  • Pratyus Patnaik

    Co-Founder & CEO · Natoma

    Proper Guardrails

    Pratyus Patnaik says organizations need proper guardrails, monitoring and auditability to stop AI agents from escaping and doing things like hacking another company, a scenario he says has become a reality.

    SiliconANGLE theCUBE · Watch on YouTube

More Cyber Security questions

View all →