Every signal in cyber right now is pointing in the same direction: AI is accelerating the threat environment, and defenders need agentic capability to keep pace.
The U.S. reversal on Anthropic’s Fabel 5 and Mythos 5 models is the latest proof point. Access was restricted over national security concerns, then lifted after Anthropic agreed to tighter safeguards and closer government coordination. That sequence captures the tension every organization is now trying to navigate. These systems are too powerful to ignore, too useful to sideline and too fast-moving to manage with conventional controls alone.
Five Eyes cyber leaders have already warned that advanced AI will reshape offensive and defensive cyber capabilities in months, not years. Models such as Mythos, Fable and China’s GLM-5.2 show why. They point to a world where complex workflows, vulnerability discovery and decision cycles can be compressed from hours of human effort into minutes or seconds of machine-led activity.
For security leaders, the question is whether their cyber defense can adapt quickly enough in the age of AI. Leaders should be asking: Is my current cyber defense strategy the right strategy for today’s environment?
AI has changed the insider risk model
Before security operations can become truly agentic, organizations need to update the risk model they are operating from. AI has dismantled the old risk model. Today, security risk forms at the convergence of humans, data and AI – both inside and outside of organizations. Before AI, insider risk was the result of negligence, malice or compromise. Now AI can amplify each of those risks, and in some cases become part of the risk itself when it has the trust and authority to act on behalf of users. AI can take an appropriate, authorized request and carry it out in a way that unintentionally harms the organization. The intent may be different, but the result is the same: damage.
It’s already happening. In one DTEX investigation, a system administrator used an agent to quietly steal data under a colleague’s login. The agent did exactly what it was asked, but pulled unsafe code from the internet to hide the theft, exposing the company to even more risk. That is just one example of many playing out across enterprises today, and why AI must be treated as part of the insider risk model.
Security teams must gain visibility into which agents exist, what they can access, and how they are being used. Then they must implement a solution to provide observability and control over those agents, regardless of whether they’re authorized or unauthorized.
The 2026 Cost of Insider Risks Global Report shows how far most organizations still have to go. Forty-four percent have minimal or no visibility into AI agent activity, and just 19 percent classify AI agents as equivalent to human insiders.
For CISOs, the near-term priority is updating their risk model, their strategy, and their tooling so AI activity is observed and managed with the same seriousness as privileged human activity.
Agentic defense is how security meets the moment
Once the risk model changes, the tooling has to change with it. To keep up with AI-driven activity, security teams need AI agents working alongside them — purpose-built to detect, interpret and respond faster and more accurately than humans alone ever could.
Security teams were already overwhelmed before AI. Now they are facing greater complexity, faster-moving threats and pressure to respond at machine speed. Agentic defense is how security meets this moment: by using defensively trained AI agents to cut through the noise, add context and help teams act with the speed and efficacy today’s environment demands.
The first test is triage. In an AI-accelerated environment, the cost of slow triage compounds quickly. Analysts cannot spend 30 minutes plus manually working out whether an alert matters. Triage is a natural fit for agentic capability because it is repetitive, context-heavy and time-sensitive. Done well, it can surface what matters, explain why and give analysts a clearer place to start. DTEX has seen that translate into real operational savings, with one government deployment saving more than 40 hours per month per analyst in alert review.
The second test is investigation. Modern incidents rarely sit in one system. They unfold across identities, endpoints, applications, files, collaboration tools and data movement. Agentic capabilities should help reconstruct that activity quickly enough for defenders to act while the risk is still forming. That means building timelines, identifying unusual behavior and showing how actions relate to intent, exposure and business impact.
The third test is control. Speed without oversight won’t be trusted, and it shouldn’t be. Any agentic security capability needs clear permissions, audit trails, explainable reasoning and human oversight for consequential actions. The goal is faster, better-supported judgment to prevent breaches before they occur or, worst-case scenario, as they are unfolding.
Behavioral context is the foundation o f agentic defense
It’s important to note that agentic defense will only be as good as the data and context feeding it. That has always been true in security, but insider risk practitioners understand it better than most. Alerts usually capture the risk event: a download, an upload, an access attempt, a policy violation. The harder and more valuable work is understanding the context and thus the intent. Understanding the intent of an action is critical when dealing with insiders and, now, AI agents.
This requires understanding intent and whether activity still aligns with a person’s role or an agent’s delegated purpose. Understanding intent requires behavioral context across who did what, with what, when, where and, most importantly, why. Without that context, AI will not create better defense. It will only move bad assumptions faster.
AI is shrinking the time between intent and impact, and Ponemon’s 2026 Cost of Insider Risks research shows the cost impact: incidents contained in less than 30 days cost US$14.2 million on average, compared with US$21.9 million when containment takes more than 90 days.
That is why behavioral context is foundational to agentic defense. A trusted behavioral dataset is what allows agents to reason accurately, move quickly and produce outcomes security teams can defend.
Looking ahead
Five Eyes leaders were right to frame this as a leadership issue, not a purely technical one. Organizations should reduce unnecessary access, strengthen identity controls, prepare for incidents and use AI deliberately to improve detection and response.
For insider risk programs, that means mapping where AI agents operate, what data they can reach, what actions they can take and what behavioral signals would indicate misuse, compromise or unsafe delegation.
AI is compressing the defender’s timeline, accelerating attacks and changing how risk forms inside the enterprise. It’s also exposing how little time defenders actually have.
The next step is not to wait for AI risk to mature, but to make it observable, accountable and defensible now. That means bringing AI agents into the insider risk model, giving analysts the behavioral context to understand what is happening, and using defensively trained agents to act before risk becomes impact.
About the Author
Marshall Heilman is the CEO of DTEX, a company specializing in AI-powered insider risk management that helps organizations prevent intentional and unintentional data breaches before they happen. With over 20 years of industry experience, he spent 17 years at Mandiant — helping grow it from startup to acquisition by Google — most recently as Global CTO. Marshall is a former red teamer and incident responder and is recognized internationally as a cybersecurity expert and speaker.
Marshall can be reached online at LinkedIn and at our company website https://www.dtex.ai/.
