AI-Enabled Offense and Defense, One Continuous Flywheel

The Market is Asking the Right Question

For two years, the industry's answer to "are we secure?" was a risk score and an attack surface map, meaning CVSS numbers and scanner dashboards. Exposure reports that took longer to read than attackers move.

That conversation wasn't wrong. It just stops short of a more important question: can an attacker get in, and what can they do once they are in? Not theoretically, or on a CVSS chart, but in your environment, against your controls, using the AI-accelerated tradecraft adversaries are running today.

This isn't a hypothetical shift. In November 2025, Anthropic publicly disrupted a China-linked actor (GTG-1002) that used Claude to execute 80–90% of a multi-stage intrusion autonomously. Three months earlier, a single threat actor (GTG-2002) used AI to extort 17 organizations across multiple industries: government, healthcare, emergency services, in one month. 

Acronis reported that by H2 2025, attackers had moved from experimenting with AI to embedding it operationally across every stage: reconnaissance, social engineering, malware development, and ransomware negotiation. The gap in the attacker vs. defender tempo chart is widening, phishing volume is up 1,265% with generative AI. Less than 25% of organizations recover within 24 hours of an AI-driven ransomware hit.

Security leaders are asking the right question now. The answer requires something the industry hasn't built: an AI-native offense and defense system that operates at the same tempo the threat actor can. eSentire recently announced our answer to this. Through our new operating model, Controlled Autonomy SecOps and AI-native offensive capabilities, we are building the connective tissue that drives rapid improvements in security posture across multiple security disciplines.

Recognizing the Problem Isn't the Same as Solving It

A real offensive capability, one that's not a scanner or a threat feed requires an adversary on your side of the table. Real tradecraft and attack simulations tied to your environment, your identity surface, your exposed assets, correlated against existing vulnerability and exposure data.

Most of the market is now looking at how to connect the dots. Some vendors are partnering or integrating. Some are pointing security teams at a separate console with a workflow that depends on two vendors agreeing on the same priorities at the same time.

Each of those approaches solves some of the problem. None of them reliably solve for: when an offensive lens on your environment determines a real attack path that delivers material business impact, what happens next?

Closed Loop or Detailed Report?

"What happens next?" may sound simple, but most security leaders don't ask it until they're three months into an engagement and wondering why the findings haven't translated into better security outcomes.

Does it push a Jira or ServiceNow ticket without context, attack path analysis, or objectives achieved? Does it generate a report that lives in a dashboard until someone has bandwidth? Or does the offensive lens drive net new detection logic, response playbooks, and hardening priorities without a handoff, multiple meetings, or two vendors pointing at each other?

The answer tells you whether you have a closed loop, or just a very detailed report.

What We've Built, and Where eSentire’s Atlas AI Lives

We've already shipped AI-native offensive security testing inside eSentire Atlas. It runs against real customer environments, generates real attack paths, and our operators are already acting on those findings. This isn't AI sprinkled on top of a legacy MDR motion. It's AI compounded across every layer of Preempt, Detect, and Respond and the connective tissue between them is what changes the outcome.

Here's what the AI is doing across the three layersf:

• On Preempt, AI agents emulate adversary tradecraft at machine speed, autonomous reconnaissance, vulnerability confirmation, attack path construction to surface exposures validated exploits that matter before attackers find them. Not annually. Persistently.
• On Detect, models triage signal, suppress noise, and correlate behavior across telemetry no analyst could hold in their head simultaneously. Mean Time to Engage (MTTE) is under 30 seconds, with agents on-signal before most teams have even opened the alert. 100% of signals are autonomously triaged, with full threat context assembled in under five minutes. The closest published industry benchmark measures detection only, not full context assembly, not engagement, not response.
• On Respond, AI drafts containment actions, accelerates investigation, and hands SOC operators a decision, not a dashboard. Customers preset response parameters; the system executes within those bounds and delivers an explainable, reversible audit trail for every action taken.

The System in Motion

Abstract architecture claims are easy to make. Here is what Controlled Autonomy SecOps looks like against a real threat, in a production environment, with a concrete outcome.

Autonomous recon mapped a customer environment and surfaced Webmin MiniServ 1.910 running on port 10000. Version matched, and reachability checks confirmed CVE-2019-15107, a critical remote code execution vulnerability, CVSS 9.8, unauthenticated. The Atlas AI agent drafted the exploit.

Then the adaptive guardrails engaged.

Engagement scope controls determined that payload delivery was outside the operational boundary. The exploit did not run. The vulnerability was surfaced, confirmed, and escalated with the full attack path documented and the finding routed directly to the customer's remediation workflow, without requiring a human to make the call to stop.

Why Home-grown Drives High-Impact Change

A partnership can share data, but it can't share decisions. It can pass a finding, but it can't share an operator. It can integrate at the API, but it can't integrate the playbook, the threat model, or the customer relationship. 

It can't share accountability when the loop breaks and something gets missed. 

The example above doesn't happen when offense and defense are different products from different vendors. The moment Atlas surfaces a CVSS 9.8 RCE and the guardrails make the call on what the AI is allowed to do next that requires the offensive engine and the governance layer to share the same threat model, owned by the same team. If that's a partnership, you're negotiating the workflow while the clock runs. 

The closed loop across Preempt, Detect, and Respond isn't a feature set you assemble. It's an operating model. One platform, one team, one place where the responsibility lives.

Why guardrails aren't optional.

Full autonomy without adaptive guardrails creates accountability gaps boards and insurers cannot accept. Three properties make AI authority defensible: explainability (a kill chain a CISO can defend, no black boxes), reversibility (every action can be undone), and policy-bounded authority (Layer 1 acts alone; Layer 2 with a human; Layer 3 with explicit consent). Without all three, you're trading speed for accountability. Controlled Autonomy requires both.

What to Ask When Offense Meets Defense

As a security leader, you're going to be asked how you combine offensive and defensive security. Here's the test that cuts through the positioning:

The right question isn't "how good is each side?" It's whether your environment got more secure this quarter. Less exposure to take advantage of fewer attack paths to exploit, more detections built, threat actors not achieving their objectives.

Ask it plainly: What got fixed? What got caught earlier? What attack stopped working? What's measurably different about your risk between last quarter and this one?

Four questions to put to any vendor claiming an integrated offense-defense capability:

• When offense finds an exposure, does defense automatically deploy a detection rule for it, or does someone have to open a ticket?
• Can your AI revoke sessions and isolate endpoints autonomously, with an explainable, reversible audit trail?
• Does the offensive simulation run on a recurring model, or only during the annual pen test window?
• Who owns the seam between detection and response:? Is it three vendors, or one operation working across all three disciplines?

If the honest answer to most of those is "we found a lot of things and talked about them", because the work crossed multiple consoles, multiple contracts, and multiple teams you don't have offense and defense working together. You have findings that generated meetings.

The model that defines this new approach is the one where every offensive finding ends in a hardened control, every attack pattern ends in a sharper detection, and your team can show the work in the language of risk reduced, not tickets opened.

ABOUT THE AUTHOR

Justin Bailey Senior Director, Product Marketing

Justin Bailey is Senior Director of Product Marketing at eSentire, where he leads go-to-market strategy for eSentire's portfolio spanning MDR, offensive security, and threat intelligence. With deep experience across multiple security disciplines, and intelligence-driven security programs, Justin specializes in translating complex security capabilities into impactful and easy to understand narratives. He works at the intersection of product, marketing, and sales to drive growth through go-to-market activities.