Since the launch of CrowdStrike AI Security Services in 2025, our Professional Services team has yet to encounter an organization with an accurate inventory of the AI tools and services in use across its environment. One customer counted 150 agents in its inventory. We found over 500. Another had not approved agentic development at all; we discovered over 70 active agents. In many cases, web filtering created a false sense of control by masking the extent of unapproved AI activity taking shape inside the environment. These are not edge cases. This is the norm for organizations of every size, across every industry and region. The new CrowdStrike Shadow AI Visibility Service aims to address this problem by giving organizations the truth about their AI footprint. Powered by the CrowdStrike Falcon® platform and delivered by CrowdStrike experts, this service uses telemetry-based evidence to identify sanctioned and unsanctioned AI usage across endpoint, cloud, and SaaS environments. In the past year, two trends have accelerated the shadow AI problem. First, many organizations have prohibited security teams from generally blocking AI tools and sites for fear of inhibiting experimentation and productivity. Second, AI adoption has accelerated, and the variety of tools has multiplied. CrowdStrike AI services engagements continue to find shadow AI in SaaS and cloud-hosted AI/ML services. We’re also finding shadow AI across the full endpoint surface: desktop AI applications, browser extensions, IDEs, packages, MCP servers, models, and frameworks. Most organizations also lack visibility into how users are interacting with AI applications, including the user prompts and LLM responses that may contain sensitive data, source code, or credentials.Shadow AI Changes the Risk Equation
Figure 1. The Falcon Adversary OverWatch threat hunting team has observed significant growth in agent-triggered detection leads, now tracking at 2.5x the rate of human-triggered leads. This demonstrates that AI agents now operate on endpoints, and they are increasingly taking autonomous and potentially risky actions. Discovering shadow AI across all of these vendors requires a security stack that sees across every surface where AI operates. Most organizations don’t have one. Shadow AI differs from traditional shadow IT because it frequently integrates into existing, approved workflows without requiring formal installation. Security teams face an immediate challenge: They cannot protect what they cannot see. And unlike shadow IT, undetected AI doesn’t just access sensitive data — it can expose this data to unauthorized systems and take autonomous action that may disrupt or jeopardize production operations. The visibility gap is driven by four primary factors.
What It Looks Like in Practice Why It Matters Stealth Access An employee uses an AI chat app via their browser or installs a computer-use agent on their machine, with no security review. Sensitive data can be exposed outside approved boundaries with as little as a prompt. Security may have no record that it happened. Agents on the endpoint inherit user permissions and capabilities that can be hijacked or misaligned. Feature Bundling AI functionality can inherit access to enterprise data and workflows before security has assessed the risk, governance model, or data handling implications. Silent Extensions A developer installs an AI-powered IDE extension, browser plug-in, or MCP-connected tool. No one in security is notified, and it begins processing code, prompts, and internal data. The organization believes it has a limited number of AI agents in production. In reality, the number is far higher, with new agents, integrations, and workflows proliferating across the environment. Security cannot govern what it cannot inventory. Untracked agents may hold excessive privileges, interact with sensitive systems, and take autonomous action that puts production operations at risk. A SaaS provider adds AI assistants or agentic features through a routine product update or license change. The tool is already approved. The new AI capability is not. These tools often sit outside traditional visibility and control points, creating risk of code exposure, data leakage, and unmonitored model interaction. Agentic Inventory Discrepancy
Without an accurate inventory, risk compounds quickly. Shadow AI is not just a funnel for sensitive data loss, including IP exposure, source code leakage, and regulatory risk. It can also act on that data by making decisions, triggering workflows, and taking autonomous action across connected systems without the visibility, guardrails, or human oversight security teams expect. This new service gives customers the evidence and guidance they need to understand their true AI footprint and reduce risk with confidence. Customers receive: A comprehensive AI inventory: A clearer accounting of AI tools, agents, copilots, extensions, and model-connected services operating across endpoint, cloud, and SaaS environments Runtime evidence of AI activity: Technical evidence of how AI is being used in practice, including prompts, responses, and agent activity — so security teams can see what’s actually happening, not what users self-report Visibility gap analysis: Analysis to understand what is present in the environment versus what the organization believes is approved — exposing unauthorized sprawl, hidden agents, and visibility blind spots Prioritized findings and expert guidance: Risk-prioritized findings and actionable recommendations from CrowdStrike experts to help teams reduce exposure and strengthen AI security posture Securing AI starts where AI executes: on the endpoint. CrowdStrike has been capturing process-level telemetry on the endpoint for over a decade, and that same visibility now extends across browser, SaaS, and cloud. This is why we can deliver AI discovery across every surface where AI operates, from a single platform and a single engagement. Visibility is the foundational phase of a secure AI strategy. Once an organization understands its real AI footprint, the next requirement is to evaluate whether those systems are resilient against adversarial manipulation. For organizations that need to go deeper, the CrowdStrike AI Systems Security Assessment extends beyond discovery into: Secure configuration: Assess feature configurations and risks in GenAI applications and services Program recommendations: Advise on governing, securing, and monitoring workforce GenAI usage and secure development of AI applications Adversarial risk assessment: Test model security, conduct threat modeling, and identify attack paths for select internally developed AI applications. Securing AI adoption requires shifting from a reactive posture to a threat-informed, evidence-driven defense. This transition begins with achieving total visibility of the current footprint. For organizations seeking greater visibility into their exposures, the CrowdStrike Frontier AI Readiness and Resilience Service provides ongoing, AI-driven scanning to identify vulnerabilities and prioritize them based on adversary risk and business criticality. As frontier AI shrinks the window between vulnerability discovery and exploitation, this helps organizations learn where they are exposed, what is reachable, and whether their controls are strong enough to stop a breach. Learn more on our services page or contact services@crowdstrike.com.CrowdStrike Shadow AI Visibility Service
Discovery Is Step One. What Comes Next?
Additional Resources
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。