The noise, confusion and fear about the impact of AI on security practices can sometimes be deafening.So, we decided to ask Priyesh Bansal, who is speaking at next month’s InfoSec World 2026 on Rethinking Zero Trust for Agentic AI Systems,” to separate some of the hype from what’s actually happening inside security teams today.Priyesh is a senior product leader at Cerebras Systems, working at the intersection of identity, security, and AI infrastructure. He currently leads identity and security initiatives at Cerebras, building a high-performance AI inference platform, and previously spent over five years at AWS working on IAM, governance, and data protection, where he launched capabilities including IAM Temporary Delegation and AWS Backup Audit Manager. He is also a co-author of a recently published article in IEEE Computer Magazine on identity and authorization controls in agentic AI systems.CRA: Where are generative and agentic AI delivering real SOC efficiency gains today, versus marketing claims?Bansal: One area where I have observed generative and agentic AI delivering real value for SOC teams is helping make sense of the sheer volume of security telemetry enterprises generate.We’ve long followed the principle of logging broadly so that, when something goes wrong, there is enough evidence to investigate and reconstruct what happened. The challenge has always been making sense of all that data. But security threats are rarely visible in a single log or event. An incident may involve a combination of unusual access patterns, lateral movement, privilege changes, and activity across multiple applications and infrastructure layers.This is where AI is helping. Rather than just replacing traditional detection rules, generative AI can help correlate signals across different systems, bring in broader context, and help analysts investigate whether something is actually suspicious and actionable in that specific environment[PB1] . Those correlations still need to be validated by analysts, however. AI can help narrow the investigation and surface what deserves attention, but human judgment is still needed to determine what actually matters in the context of that organization.CRA: Where has AI proved ineffective or even counterproductive?Bansal: AI can become counterproductive when we move too quickly from detection to autonomous remediation.An agent may correctly identify an anomaly but still lack the business and operational context to understand the consequences of acting on it. In practice, I have seen that the remediation action an agent recommends is not always the best course of action and even when it is sound, it often cannot be executed immediately[PB2] . A change in a production environment, for instance, usually needs pre-work with the engineering team to understand dependencies and sequence it safely, so the fix doesn't cause more disruption than the original issue.CRA: What new skills and roles are required in security teams as AI absorbs Tier-1 triage work?Bansal: The honest answer is that most of these skills don't disappear as AI absorbs Tier-1 work - they move up a level. As the repetitive triage gets automated, the value of a security analyst shifts from processing alerts to exercising the judgment AI still can't. I'd group the skills that matter into two.One is a skill teams have always needed but can now lean on far more heavily: judging the risk and impact of something in the context of the specific business. An agent can gather and connect a lot of information about what happened, but deciding whether it actually matters, what the downstream consequences of acting are, and what is safe to automate versus what needs a human that depends on understanding the business. As AI handles more of the mechanical work, this judgment becomes the differentiated human contribution.The other skill is genuinely new: building real expertise in how these agents and tools actually work, and what new threats they introduce, so the right safeguards can be put in place. Teams now need people who understand how an agent accesses data, takes actions, and can itself be manipulated or compromised — and who can turn that understanding into concrete controls[PB3] . I don't see a new job title so much as existing roles shifting toward setting guardrails and deciding where agents can act independently versus where humans stay involved.CRA: How do you govern AI tool adoption inside security operations without creating new blind spots?Bansal: The more context agents have and the more deeply they are embedded in your ecosystem, the more useful they become. An agent that understands your internal data, applications, workflows, and systems can provide far more value than a standalone chatbot.The goal should not be to give an agent broad access just because it becomes more useful that way. The challenge is figuring out how much context and authority it actually needs for a specific task and keeping that access appropriately scoped.The problem becomes [PB4] even more complex when AI starts taking actions and especially when workflows span multiple agents. Identity, authorization scope, and context can be lost as a request moves across agents and systems, making it difficult to determine who acted and under whose authority. It becomes important to distinguish between what a human did directly, what an agent did on behalf of a human, and what an autonomous agent did within pre-authorized boundaries.CRA: What dimensions should security teams consider when evaluating the risk of an AI agent?Bansal: I use three dimensions to evaluate the risk:
The higher the potential impact, the stronger the permission boundaries, controls, and oversight should be.
We also need to start evaluating agents themselves as first-class actors, similar to how we evaluate human users and other non-human identities. If an agent takes an action, we should be able to attribute it to that agent specifically. And just because a user has access to something does not mean an agent acting on their behalf needs the user's entire permission set — the authority given to the agent should be scoped to the specific task and remain intact as the request moves across multiple agents and systems.CRA: What is the appropriate and realistic role of the cybersecurity team when it comes to adoption of AI tools throughout the rest of the organization? Is the answer different for enterprise tools vs. apps adopted by individual users?Bansal: The role of cybersecurity should not be to centrally approve every AI tool or use case. A more realistic role is to set the baseline guardrails, define the risk framework, and provide reusable security controls, so the organization can adopt AI without reinventing the security model for every use case.There is a real difference between enterprise tools and individual apps adopted by employees:
Ultimately, security teams need to shift from primarily asking, “Is this AI tool approved?” to asking more specific questions: What can the system access? What is it authorized to do, and under whose authority? How could it itself be compromised or manipulated? That shift from a one-time binary approval to an ongoing, risk-based evaluation is what lets the organization adopt AI quickly without losing control of it.
- Action sensitivity — what can the agent actually do;
- Data sensitivity — what information can it access; and
- Execution environment and exposure — how broadly connected or exposed is it.
We also need to start evaluating agents themselves as first-class actors, similar to how we evaluate human users and other non-human identities. If an agent takes an action, we should be able to attribute it to that agent specifically. And just because a user has access to something does not mean an agent acting on their behalf needs the user's entire permission set — the authority given to the agent should be scoped to the specific task and remain intact as the request moves across multiple agents and systems.CRA: What is the appropriate and realistic role of the cybersecurity team when it comes to adoption of AI tools throughout the rest of the organization? Is the answer different for enterprise tools vs. apps adopted by individual users?Bansal: The role of cybersecurity should not be to centrally approve every AI tool or use case. A more realistic role is to set the baseline guardrails, define the risk framework, and provide reusable security controls, so the organization can adopt AI without reinventing the security model for every use case.There is a real difference between enterprise tools and individual apps adopted by employees:
- Enterprise tools that are deeply integrated into company data and systems need stronger architectural and governance controls, because they take on much broader context and can have a much larger blast radius.
- Individual apps start out as more of a data-handling and acceptable-use question. My view is that these should generally be restricted from connecting to enterprise applications or systems. That connection is the point where a personal-productivity tool quietly becomes an enterprise-integration problem, with far broader access and impact. The residual risk is mostly about what information employees share with these tools, and that can be addressed through employee training rather than trying to block every app[PB5] . The goal of that training should be to enable employees to use these tools safely by helping them understand what information is appropriate to share, where the risks are, and when additional security guidance is needed. Some of that judgment ultimately has to sit with the people using the tools day to day.
