> ## Content Index > Fetch the complete content index at: https://www.cybrsecmedia.com/llms.txt > Use this file to discover other available public pages before exploring further. # BNSF CISO Alex Levy on Agentic AI Risk: 'Baby-Proof Your Environment' - URL: https://www.cybrsecmedia.com/bnsf-ciso-alex-levy-on-agentic-ai-risk-baby-proof-your-environment/ - Published: 2026-11-11T13:03:37.000Z - Updated: 2026-11-11T13:03:57.000Z - Description: Alex Levy on ring-fencing agents, celebrating mistakes, and recovering in minutes. - Author: George V. Hulme - Tags: Agentic AI, Commvault Shift, Article The next destructive event in your environment may not come from an attacker. It may come from an AI agent doing its best to please you. That is the working assumption of Alex Levy, chief information security officer at BNSF Railway, who was named the inaugural TIME and Commvault CISO of the Year at Commvault's Shift conference this week in Nashville. Levy is responsible for the security of a rail network spanning 33,000 miles, and part of that job is now deciding how much autonomy to give software that can act like a child. "Their goal is to please," Levy said of AI agents during an on-stage interview with TIME CEO Jessica Sibley. "Even if you give them an instruction, they're not necessarily going to follow it to a T. They're going to act like a child almost and go and do what they want to do. So you have to protect yourself. You have to baby-proof your environment," he added. [Subscribe to the CYBR.SEC.Media newsletter](#/portal/signup/free) The metaphor carries into a broader argument from Levy: the security basics enterprises let slide for years are required for a reliable control plane for agentic AI. Identity management, which Levy said "nobody does well," tops his list. Non-human identity was a hard problem before agents arrived. Now a non-human actor performs potentially destructive actions on behalf of a human, compounding every over-provisioned account and permissive firewall rule already in the environment, he said. "There are firewalls that look like Swiss cheese. There's identity with users that are significantly over-provisioned," Levy said. "All of that needs to be the focus of where you start. The only way you're going to be able to ring-fence these agents is by having a good handle on what they can do and when." ## Chaos to the left, chaos to the right For decades, Levy said, humans were the chaos in the security equation and computer systems, and software were predictable: configure a server correctly and you knew how it would behave. AI removes that predictability. "Now you have chaos on both sides of the equation. The system is no longer deterministic," he said. Levy stressed agentic AI security success is also organizational. Levy owns not just cybersecurity at BNSF but the broader technology risk function: business continuity, disaster recovery, governance, risk, and compliance, and he argues the most important control is cultural: an environment where employees report their own mistakes quickly. "Celebrate the error," he said. "When somebody makes a mistake, don't villainize it, don't demonize it. Do so in a public way. You're going to find that you're going to identify more issues in your environment that maybe your observability didn't catch." With Levy citing an incredible volume of vulnerability discoveries per month, he considers it "fundamentally impossible" to keep up. Detection and human reporting must cover what prevention cannot. That also means security teams can no longer stay within their traditional role. "We can't be the org of no anymore," he said. "AI is going to push its way through whether we want it or not." The goal instead is what he called a healthy amount of friction: enough that users can "stub their toe" without creating untenable risk. He offered one thought experiment for boards measuring their own exposure: if your organization has leaned into AI-assisted coding for three to six months, what happens to productivity if those tools go dark for a day? "If your developers who've been developing for 20 years suddenly lose their ability to code because the tool they've gotten so accustomed to is not available, that's a huge business risk." ## Old systems, attackers with new tools Levy also noted the newfound risks associated with vulnerabilities within older technology. Defenders have long assumed aging systems, such as OT estates, carried a kind of incidental protection: the expertise to attack 15-year-old controllers and 20-year-old wayside switches had aged along with the systems. AI, he said, has ended that. "What's old is new again," Levy said. "AI is now giving mainframe capabilities to people who have never touched them before. It's created a new threat vector we all need to be cognizant of,” he said. When asked what every CEO should ask their security chief, Levy's answered: observability. How do you know something has gone wrong before your customers call to tell you? Defenders must also be cognizant of how quickly attackers can now operate. Environments that once took attackers days to dismantle can now disappear in minutes, and recovery time objectives measured in days no longer hold. "It needs to be minutes, and in the worst-case scenario, hours." The takeaway is one that’s been thirty years in the making. The industry must move from preventing every failure to surviving any one failure. **Related:** [AI Agent Damage Takes Seconds, Recovery Takes WeeksAI agents can wreak havoc in seconds, but business recovery can take weeks. New Commvault tools aim to close that gap by restoring applications, identities and agentic AI workflows.![](https://storage.ghost.io/c/ab/67/ab676516-71e3-473d-8f73-9e0692f5aaee/content/images/icon/CYBR.SEC.Media-Logo-copy-28eac616-b52a-46bb-a1ea-e4c5722cd1a4.jpg)CYBR.SEC.MediaGeorge V. Hulme![](https://storage.ghost.io/c/ab/67/ab676516-71e3-473d-8f73-9e0692f5aaee/content/images/thumbnail/3ca19142-f4c3-418f-8896-1f012143acbc-9b121e54-349d-489f-8a4b-c9e6d528973a.png)](https://www.cybrsecmedia.com/ai-agent-damage-takes-seconds-recovery-takes-weeks/) [![](https://storage.ghost.io/c/ab/67/ab676516-71e3-473d-8f73-9e0692f5aaee/content/images/2026/11/2027-Events-Banner-27.png)](https://www.cybrsecmedia.com/conference/)