If your next cybersecurity budget conversation begins with whether AI makes the existing security stack obsolete, stop. That is the wrong decision frame and a fine way to turn a serious planning exercise into a vendor slogan contest.

Two events in 2026 have made the confusion understandable. Anthropic’s Mythos capability showed that an autonomous agent could find exploitable weaknesses at machine speed. A later intrusion at Hugging Face showed autonomous agents moving through permitted infrastructure and reaching customer datasets while pursuing evaluation objectives. One was a controlled research capability. The other was a real operating account. Neither says enterprises can stop spending on security.

Together, they say something more useful: the value inside cybersecurity is being repriced. The executive question is not whether network, endpoint, identity, cloud, security operations, exposure management, application security and recovery still matter. They do. The question is which controls deserve the next dollar and what evidence should be required before approving it.

The Security Market is Not Collapsing

It is tempting to treat every AI breakthrough as a zero-sum event in which a new model destroys a market overnight. The operating evidence does not support that story in cybersecurity.

CrowdStrike reported quarterly revenue of $1.47 billion, up 26%, and annual recurring revenue of $5.84 billion, up 25%, in its August 2026 results. Palo Alto Networks reported 34% quarterly revenue growth and 63% growth in next-generation security ARR, although acquisitions changed the reporting perimeter and make a simple year-over-year comparison misleading. Zscaler reported full-year revenue and ARR growth of 25%, with ARR growth of 20% excluding Red Canary. Commvault reported fiscal-year revenue growth of 19%, ARR growth of 21% and SaaS revenue growth of 52%.

Those figures do not prove that every incumbent will win or that agent security drove the results. They do show continuing demand across major security and resilience businesses. They also illustrate why executives should read the labels. Growth, acquisition contribution, cash generation and investment intensity answer different questions. A funding valuation is not a security-effectiveness test, and a strong category does not guarantee strong economics for every vendor in it.

Four Outcomes Deserve a Larger Share of Attention

The first is accountable authority. Enterprises are about to create far more machine actors with the ability to call tools, obtain credentials and make changes. An agent must have its own identity, a relationship to a human or organizational owner, a reason for its authority, a time limit and a reliable revocation path. Authenticating the initial request is not enough.

The second is enforceable boundaries. Executives do not need to become network architects, but they should insist that teams can explain what an agent may reach after it has been authorized. Identity, runtime permissions, workload isolation, network destinations and data access are different boundaries. If the plan depends on every layer behaving perfectly, it is not much of a plan.

The third is verified exposure reduction. Faster vulnerability discovery does not magically create more engineering capacity. It can just produce a larger backlog with better AI-generated prose. The value lies in validating which paths matter, assigning ownership, making a safe change and proving the exposure was reduced. Counting findings is not the same as reducing risk.

The fourth is trustworthy recovery. Autonomous systems can corrupt data, propagate bad changes or trigger business actions that cannot be undone by restoring a server. A database may be recoverable while a disclosed secret is not. A payment may require a compensating transaction. An agent’s restored memory may recreate the condition that caused the problem. Recovery plans have to cover trusted business state, not just infrastructure availability.

Platforms and Specialists Both Have a Case

The next 24 to 36 months will not produce a clean platform-versus-startup winner. Large platforms have installed control points, distribution and the ability to connect identity, network, execution and response. Specialists can win where customers need a control that crosses competing environments or where a new workflow is too important to remain a bundled checkbox.

The capital flowing to companies such as Tines, Chainguard, Noma, Torq, XBOW, Socket, Runlayer and Horizon3.ai began before Mythos and continued after it. That matters because it punctures the idea that one April announcement created the category. It does not prove that every funded vendor has a durable business. Capital establishes that investors see an opportunity. Renewals, expansion, production outcomes and the amount of human supervision required establish whether the opportunity became a company.

A hybrid market is the most plausible result. Some controls naturally belong close to identity, network and execution infrastructure. Others benefit from independence across those platforms. Most enterprises will continue operating mixed estates no matter how many keynote slides promise a unified future.

What to Require Before Approving the Next Purchase

Start by retaining effective controls. Do not remove a capable endpoint, identity or network platform because a startup has a sharper agentic story. At the same time, do not assume an old contract covers a new execution path. Extend inventories to agents, their owners, the tools they can invoke, delegated credentials and the business resources within reach.

Then insist on workflow-level testing. Procurement should ask how the control behaves after authority is delegated, when credentials are cached, when a service becomes unavailable and when the model provider changes. Reference customers should be able to describe how long deployment took, how many exceptions remain, what supervision is required and what operating result improved.

Finally, separate technical claims from economic claims. If a vendor says it reduced reachable value at risk, ask how assets were valued, which dependencies were included and whether correlated losses were double-counted. Reachability can be a powerful executive lens, but reachability is not expected financial loss, and a marketing estimate is not a customer-verified saving.

The cybersecurity budget is not disappearing. The tolerance for products that observe a problem and leave the consequential work to someone else should. That is the repricing leaders should bring into the next planning cycle.

Download the Full Report