Enterprise defenses hunt the unauthorized: the foreign file, the stolen credential, the behavior nobody sanctioned. Agentic AI security inverts that model, because the agent moving through the business was invited in and granted standing access to critical systems
That inversion has opened a market for startups built around the runtime behavior of autonomous software. Traditional threat models assume an attacker is an outsider, but an agent operating with delegated authority looks legitimate right up until it does not, according to Amol Mathur (pictured, right), co-founder and chief product officer of Dash Security Inc.
“If you think about AI agents, they are fundamentally different as a technology. They are non-deterministic, number one. They mix data, instructions and the input that they get,” Mathur said. “Unlike other technologies, you are giving the AI agents a lot of delegated authority, which means that they are supposed to have access to a lot of systems and critical data.”
Mathur and Maor Hod (left), co-founder and chief executive officer of Dash Security, spoke with theCUBE’s Dave Vellante and Rebecca Knight at the Fal.Con event, during an exclusive broadcast on theCUBE, SiliconANGLE Media’s livestreaming studio. They discussed why agent behavior must be governed at runtime and how intent shapes detection. (* Disclosure below.)
Agentic AI security begins with visibility and runtime control
Fragmented tooling is the first obstacle. Sanctioned tools sit alongside shadow deployments on workstations, in browsers and across cloud environments, leaving teams with no clear view of how data crosses trust boundaries. Dash maps that AI estate before applying policy.
“First we start with visibility, providing the right visibility to understand what people are doing with AI, with what AI tools they have within their organization,” Hod said. “Then to give the right size guardrails in order not just to help customers to block stuff, also to make sure that they can adopt AI in confidence, but at the same time not slow the organization.”
That context turns blunt controls into precise ones. One customer with 140,000 employees was blocking users whose reports contained nine-digit numbers that had nothing to do with Social Security records, Mathur noted.
“We understand the agent’s role and their past history,” Mathur said. “We can say that this is a guy in operations who’s running a report on something which has a nine-digit number, which has nothing to do with Social Security. We were able to reduce false positives by 95%, and they rolled it out with action automatically being taken for their entire user base.”
Dash integrates with CrowdStrike Holdings Inc., pulling endpoint telemetry into the agent session and feeding the stitched signal to Falcon Next-Gen SIEM, CrowdStrike’s security data and detection platform. For a company founded in October 2025, agentic AI security means tracking the frontier labs as closely as the attackers.
“Security follows the curve of the technology that you’re protecting,” Mathur said. “One of the core focuses we always have is to deeply understand on one hand how generative AI and agentic AI [are] changing and evolving, what the frontier labs are doing. The second thing is [to] deeply try to understand how our customers are adopting and using this technology.”
Here’s the complete video interview, part of SiliconANGLE’s and theCUBE’s coverage of the Fal.Con event:
(* Disclosure: TheCUBE is a paid media partner for the Fal.Con event. Neither CrowdStrike, the sponsor of theCUBE’s event coverage, nor other sponsors have editorial control over content on theCUBE or SiliconANGLE.)
