Agentic AI security: attack surface shifts to machine speed

Agentic AI is not introducing a fundamentally new attack surface, but it is compressing response time from human-scale to machine-scale. CrowdStrike and Box are among the first to rebuild detection and governance around that reality, and early adopters are already discovering gaps they did not anticipate.

What CrowdStrike and Box are doing differently

CrowdStrike extended the Falcon platform to police AI agents at the endpoint, treating each agent as an asset with an identity and a data footprint attached. According to Cristian Rodriguez, field CTO at CrowdStrike, every enterprise asset has a type of system, an identity, and a type of data it can access, and AI automates and accelerates that experience from start to finish. Without the right guardrails, autonomous agents can do a lot of damage.

Box, which added AI agent governance controls in July, now inherits CrowdStrike’s device posture score to decide whether a content access request is sanctioned. Heather Ceylan, Box’s CISO, explained that attack surface is the same, but agents move at machine speed, so detections and visibility must be real-time.

Early adopters who deployed agents six to twelve months ago are now returning to CrowdStrike asking for visibility and data controls across SaaS apps, endpoints, and cloud instances. As Rodriguez put it, "the AI sprawl is real" and enterprises have "bitten off a little more than they can chew."

Why this matters for AI builders

If you are shipping AI agents that read, write, or move enterprise content, you are inheriting a security problem that has no settled architecture. There is no shared responsibility model for agentic AI the way cloud computing eventually developed one. Ceylan predicted it will take two to three years before the industry agrees on what a secure AI architecture looks like.

That uncertainty matters for product decisions. If your agent relies on SaaS APIs, endpoint access, or cloud infrastructure, you need to plan for how those systems will enforce identity and data boundaries. The platforms you build on today may change their posture controls, and your agent could be blocked or permitted based on device-level signals it does not control.

What changes in practice

Security teams now need real-time visibility across three domains: endpoints, SaaS applications, and cloud instances. Traditional periodic scans or human-in-the-loop reviews will not keep up with agents that can execute tool calls in milliseconds.

Governance programs must also scale. Enterprises that treated early agent deployments as experiments are now finding they need formal policies around what data an agent can access, what tools it can call, and how its identity is tracked across systems. CrowdStrike’s approach of treating each agent as an asset with an identity-bound data footprint offers one template, but it is far from universal.

The unsettled parts

The evidence for these patterns comes primarily from SiliconANGLE’s coverage of Fal.Con and related events. Concrete security patterns for agentic AI are still developing, and no single vendor has a complete solution. Pricing, deployment constraints, and responsibility models remain organization-specific. Builders should watch how CrowdStrike, Box, and others evolve their approaches, but expect to build internal governance layers until the industry standardizes.