AI has widened the attack surface from both directions, leaving defenders with two problems to solve at once: agentic attacks coming at them, and the increasing AI usage within their own organizations.
On the first front, AI handed attackers three structural advantages at once: speed, scale, and the removal of the skills barrier. Full attack cycles that took weeks of tradecraft now run in minutes, with the fastest data exfiltration of 2025 clocking in at 6 minutes. What once required an army, and deep expertise, now runs on a single prompt for anyone with the right model.
On the second front, the surface expands from the inside. Enterprise data already lives everywhere, and it grows with every new business unit, location, and piece of software the company adopts, especially AI software. Every model, agent, and assistant a team turns on becomes another asset to secure. Adopt fast without securing it, and the risk comes from inside: data leaking into public models, agents acting beyond the scope they were given, and sensitive information exposed to people who were never meant to see it.
Two fronts, one surface. Neither the pace of agentic attacks nor the rate of internal AI adoption slows down for a human-speed defense. Follow these five steps to build an agentic defense that protects the entire environment at the speed both fronts move:
Step 1: Close the Machine-Speed Gap With Automation
Attackers used to spend weeks on resource development and days on initial access. With jailbroken models and off-the-shelf exploit kits selling for less than $150, a low-skill actor can spin up a targeted campaign in minutes, point an agent at a company, and let it run recon to find the crack in the armor. Dwell time that used to run 6 to 9 months is now closer to 10 minutes before execution.
Adding analysts to machine-speed attacks slows response down. Set the rule accordingly: nothing should live in the environment longer than 30 minutes. Anything that does means containment was not automated, or the threat was not identified fast enough. Teams that automate containment and remediation have gone from 50-person manual SOCs to roughly 10, shifting the rest into threat hunting and vulnerability management, and have dropped detect-and-respond times from days and hours to minutes and seconds.
Step 2: Treat the AI Your Business Adopts as Part of the Attack Surface
Defending against AI-accelerated attacks is only half the job. 75% of employees using unapproved AI have shared sensitive data with it. Gartner predicts that by 2030 more than 40% of enterprises will experience security or compliance incidents linked to unauthorized shadow AI.
The risks are concrete. Something as simple as a common misconfiguration, where a file shared by link in Microsoft 365 becomes readable across the whole organization, lets an AI assistant connected to that environment surface data a user was never meant to see. Employees blocked from uploading to a public model will photograph a slide and send it anyway. The productive answer is enablement with guardrails: give people the tools they want, then use the compliance APIs now available to gain real-time visibility into what they feed those models and block what crosses the line.
Step 3: Scope Every Agent's Permissions, Then Validate That You Can Trust It
As you deploy AI to defend, govern what it can touch. An agent should act with the access of the person directing it, not inherit super-user rights. Scope every tool it can call line by line: an agent that can isolate a host should not automatically be able to pull files from it. Take each application on its own merits, decide what data it sees, and confirm whether it acts as its own persona or borrows an admin's. Get this wrong and tracing how data was exfiltrated becomes a nightmare.
Scoping permissions sets the boundaries, but boundaries alone will not earn a leadership team's trust to let AI drive daily operations. That trust comes from continuous testing and validation, running before, during, and after deployment rather than as a one-time pre-launch check. AI needs layered guardrails against prompt injection and scope expansion, objective quality scoring, and golden datasets that catch performance drift before it reaches production.
Step 4: Test Your Defenses Like an Attacker
Continuous validation ties the program together. The most effective approach mirrors what an adversary does: map every attack path across your environment, tracing each route an attacker could take through your identities, permissions, exposures, and controls before one ever shows up. A comprehensive solution can launch a campaign in plain language instead of scripting it. Point it at a single overly privileged identity, emulate a named threat actor targeting your industry, or replay a recent pen test to see which controls still hold.
Mapping the path is not enough on its own. When you need proof, execute the technique against the live environment to confirm what is real, then feed every finding back into the loop as new detections, hunts, playbooks, and remediation. That closed cycle turns validation from a point-in-time audit into something continuous:
Attack-path mapping across identities, permissions, and exposures, refreshed as the environment changes
Proof through execution against the live environment
Findings that harden the SOC automatically, building detections and hunts for paths already in use
Make it run continuously, safely, and without runaway cost, so any defender can pressure-test the environment as fast as attackers are probing it.
Step 5: Build Governance That Moves at the Speed of the Business
None of this holds without governance that keeps pace. Avoid the two failure modes: the committee that accepts every new tool without thinking, and the one that says no to everything. Sit in the middle, get visibility into what AI tools teams are adopting, wire security into the procurement chain so unsanctioned purchases surface early, and partner with the business instead of blocking it. Bring business, technical, and revenue leaders to one table so the organization buys AI safely and at scale. Move fast, because your users will build their own plan before a slow committee does.
The throughline across all 5 steps is the same: you defend AI with AI. Start with your top 5 alerts, automate them, and go again.
