CrowdStrike has unveiled a security tool that attacks its own customers on purpose. Announced on November 4 at Fal.Con Europe in Barcelona, Agentic Adversary Emulation is designed to behave like an autonomous, AI-driven attacker inside a customer's environment, chaining together vulnerabilities, misconfigurations and identity weaknesses until it finds a workable path to that organisation's most critical assets.
The new offering is powered by SafeMind, CrowdStrike's family of security models and execution frameworks, and it runs on the company's Falcon platform. Instead of producing another list of isolated flaws to patch, the system is meant to show which combinations of weaknesses actually lead somewhere dangerous, and in what order an attacker would exploit them.
An attacker that thinks in chains, not checklists
At the centre of the tool is Red Tempest, SafeMind's offensive model, which operates like an autonomous attacker. CrowdStrike says SafeMind is trained on Falcon sensor telemetry, expert-labelled threat intelligence and the frontline experience of its analysts and incident responders, grounded in intelligence on more than 290 tracked adversaries, emerging agentic threats and newly disclosed vulnerabilities.
The company points to a stark example from its 2026 Agentic AI Campaign report: an AI agent chained a single vulnerability into cluster-administrator access and then exfiltrated 726 databases, with no human at the keyboard. That, CrowdStrike argues, is the threat model defenders now have to test against. A single unpatched flaw may look minor on its own, but in an agentic attack it can be the first link in a chain that ends at the crown jewels.
Agentic Adversary Emulation therefore automates the full loop: attack, prioritised remediation, and validation. After fixes are applied, it re-runs the attack to confirm that each path is genuinely closed, rather than assuming a patch did the job.
Early access only, for now
The capability is not yet broadly available. CrowdStrike says it is in early access for select customers, offered through partners in its Project QuiltWorks programme. That limited rollout means independent results and wider pricing details are still to come, and the company notes that unreleased features remain subject to change.
For the customers in that first wave, the pitch is a shift from periodic, point-in-time testing to continuous, adversary-style validation: the same environment, tested the way a real attacker would test it, with the results ranked by what actually puts critical assets at risk.
Why this matters beyond one vendor
The launch also signals a broader change in cybersecurity. As attackers begin to use AI agents that can explore, adapt and chain exploits without constant human direction, defenders are answering with AI that mimics the offence. The question is no longer just whether a vulnerability exists, but whether an autonomous attacker could string several small weaknesses together into a breach.
If Agentic Adversary Emulation works as described, it could help security teams spend less time chasing low-risk alerts and more time closing the handful of paths that matter. If it does not, it will at least have forced a useful conversation: in an agentic era, a security posture is only as strong as the weakest chain an AI attacker can find.



