AI Security
getty
A decade ago, enterprises hesitated to move critical workloads to the cloud. The technology was capable, but governance wasn’t. Security vendors that positioned themselves as the control layer captured enormous market value, CrowdStrike among them. The company's Falcon platform rode the cloud transition, becoming a dominant force in enterprise security. Now CrowdStrike is making the same bet on AI, and the evidence suggests it has a head start.
The threat environment has shifted in ways that make that bet urgent. CrowdStrike CEO George Kurtz disclosed at the recent RSA Conference 2026 in San Francisco that the fastest recorded adversary breakout time, the window between initial access and lateral movement inside a network, has dropped to 27 seconds, with the average now at 29 minutes, down from 48 minutes in 2024. The company's 2026 Global Threat Report found an 89 percent year-over-year increase in attacks in which adversaries deployed AI. These are measurements of what is already happening, not projections.
Michael Sentonas, CrowdStrike's president, puts the consequences of frontier AI models in blunt terms: the cost equation for attackers has changed permanently. "You do not need any capability," Sentonas told me in a recent briefing. "You just need a prompt and intent." The democratization of sophisticated attack capabilities has now reached its logical endpoint. Every organization, regardless of size or sector, faces the possibility of encountering a zero-day-class threat on any given day.
CrowdStrike’s response to this environment is an expansion of its Falcon platform across the full AI security stack, announced at RSAC 2026. Its strategy treats the enterprise endpoint as the control plane for AI governance. This is the place where AI agents ultimately execute, data flows, and malicious or unauthorized behavior can be intercepted in real time.
CrowdStrike’s platform covers several interconnected problems:
The architectural logic is consistent across all these capabilities: because AI agents run on endpoints, access data via identity credentials, and trigger downstream workflows, securing them requires sensor-level visibility at the device level. This is the same visibility that CrowdStrike built its business on.
Sentonas frames the agent problem precisely, saying, "you're not just securing software. You're securing autonomous behaviors." He went on to say that traditional perimeter and application controls were not built for systems that are goal-seeking by design.
The most significant recent development in AI-driven cybersecurity did not come from a security vendor, but from the frontier model builders themselves. Both Anthropic and OpenAI have concluded that the offensive capabilities of their most advanced models now require a structured, restricted approach to deployment. Both have also launched formal programs to put those capabilities in defenders' hands first.
Anthropic launched Project Glasswing alongside a limited preview of Claude Mythos, its most capable model to date. The trigger was straightforward. During internal testing, Mythos demonstrated the ability to autonomously identify and exploit zero-day vulnerabilities across all major operating systems and web browsers.
Faced with a model that surpasses even the most skilled human security researchers at finding exploitable flaws, Anthropic chose not to release it publicly. Instead, it assembled a coalition of twelve founding partners, committed up to $100 million in usage credits, and structured access with a binding requirement that the model be used exclusively for defensive purposes.
The founding partner list for Project Glasswing reads like a cross-section of the organizations responsible for the largest portion of the world's shared attack surface, including Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks.
The presence of both CrowdStrike and Palo Alto, the two largest independent cybersecurity vendors, alongside hyperscalers and critical infrastructure operators, highlights Anthropic's intent to bring adversarial expertise, not just compute capability, into the feedback loop.
Sentonas described CrowdStrike's role as a closed-loop arrangement in which the company contributes real-world intelligence gathered by tracking 280-plus adversary groups daily, evaluates how the model can be used offensively, and provides feedback on the guardrails required before broader deployment.
Days later, OpenAI followed with a parallel but structurally distinct approach. The company unveiled GPT-5.4-Cyber, a variant of its latest flagship model, fine-tuned for defensive security work, including binary reverse-engineering capabilities that enable analysis of compiled software without access to the source code. It also expanded its Trusted Access for Cyber program, introduced in early 2026.
While Anthropic restricted Mythos to 12 named partners plus roughly 40 additional vetted organizations, OpenAI opted for a tiered identity-verification model designed to scale to thousands of individual defenders and hundreds of security teams.
The philosophical difference is meaningful. Anthropic controls access through organizational vetting and contractual obligations, while OpenAI controls access through identity verification and know-your-customer processes, with the goal of making the capability as broadly available as responsible practice allows. CrowdStrike is also integrated into the OpenAI ecosystem, and the Falcon platform's AIDR capability already works with OpenAI's models.
The two initiatives converge on the core idea that the capabilities now embedded in frontier models are too dangerous to release without governance structures, and that security vendors are best positioned to build them.
That philosophy elevates the strategic importance of firms in both programs. Palo Alto Networks participates in Project Glasswing, and CrowdStrike participates in both programs. Security vendors outside these coalitions face a growing capability gap that market position alone cannot close.
CrowdStrike is not alone in pursuing AI security as a platform expansion. Palo Alto Networks announced Prisma AIRS 3.0 at RSAC, extending its cloud security platform to AI agents with runtime controls and an agentic identity provider. Microsoft has integrated AI security into its Defender and Sentinel stack, and the combination of Microsoft Sentinel and Security Copilot offers a credible alternative architecture for organizations already standardized on Microsoft. Startups such as HiddenLayer and Protect AI focus specifically on machine learning model security, a narrower but technically deep segment of the problem.
CrowdStrike's differentiation rests on three factors that are difficult to replicate quickly:
For enterprises evaluating AI security vendors, the Glasswing and TAC partnerships serve as useful proxies for assessing depth. Participation is not self-selected, as both Anthropic and OpenAI chose their partners based on their ability to contribute adversarial expertise and endpoint coverage, rather than on commercial relationships. Vendors in those programs have access to model capabilities and vulnerability intelligence that others do not. That gap widens as frontier models continue to improve.
CrowdStrike's financial results reflect what Sentonas describes as the largest driver of security demand since enterprise computing moved to the cloud. The company has reported two consecutive quarters of reacceleration in its core endpoint business, which it attributes directly to enterprise concerns about whether legacy security tooling can handle AI-era threats. Falcon Next-Gen SIEM grew 75 percent year-over-year as of the company's Q4 FY26 earnings.
These numbers reflect a buyer dynamic in which enterprises are upgrading because they believe their current posture is inadequate, and that assessment is narrowing as AI agent deployments move from pilot to production.
The urgency is also evident in the broader regulatory environment. The EU AI Act's next enforcement phase takes effect later this year, imposing cybersecurity requirements on high-risk AI systems and incident-reporting obligations, with penalties of up to 3% of global revenue.
Enterprises deploying AI in regulated industries now need the technical controls to demonstrate compliance, not merely the intent to provide them. CrowdStrike's AIDR capability, with its audit trail, prompt visibility, and policy enforcement, directly addresses that requirement. For organizations that have deferred AI governance investment, the window to do so quietly is closing.
The coordinated moves by Anthropic and OpenAI to gate their most capable cyber models signal something broader than product strategy. Both companies have concluded that frontier AI capability and cybersecurity are now inseparable. This means the most powerful models can’t be deployed without involving the firms that understand how those capabilities will be weaponized.
CrowdStrike's argument is not that it has solved the AI security problem. Rather, it is that it built the platform, the telemetry base, and the adversarial intelligence over a decade, making it the right partner as that problem scales. The frontier model builders appear to agree.
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。