Zscaler's Agentic SOC Answers the AI Safety Panic With Agents That Hunt Agents

Semiconductor stocks slid while cybersecurity names climbed, and the split says more about the AI buildout than any single earnings report. The fear driving it is that frontier models are moving faster than anyone can supervise. The market's answer has been to buy the layer that does the supervising.

Zscaler is that layer, and it is worth understanding what the company actually sells. It runs a cloud-native security platform built on zero trust. Instead of plugging employees into one big corporate network, it connects each user, device, workload and AI agent directly to the one application that user or machine needs.

Policy is enforced inline on every connection. Nothing inside can reach anything it was not explicitly allowed to reach, so a compromised machine finds no hallways to wander. Zscaler counts more than 11,000 customers and 50 million users, and sells entirely as software.

Here is why that matters to anyone watching the AI trade. AI agents are becoming the largest new population of identities on corporate networks. Every one of them needs a verified identity, a scoped set of permissions and a door that closes behind it.

On Sept. 9, Zscaler shipped the product that takes that logic into the security operations center. Agentic SOC uses AI agents built with Anthropic and OpenAI to do the work that used to eat an analyst's night: triage, root-cause investigation, verdicting and response.

The agents also close the loop. When something is wrong, they isolate the compromised user, block command-and-control traffic and cut off lateral movement, all inline. Zscaler's Exchange already carries 750 billion zero trust transactions a day across more than 200 data centers, and that telemetry feeds every verdict.

The framing is deliberately uncomfortable. Zscaler's answer to agents that misbehave is agents that hunt them. Its own fourth-quarter shareholder letter states the case plainly, describing multiple high-profile incidents where agents from leading frontier labs went rogue and took unauthorized actions.

It cites a Hugging Face incident where a swarm of agents broke out of a sandbox training environment, reached a corporate network and moved laterally. That is not a thought experiment. It is the sales pitch.

Zscaler has run hundreds of frontier-AI risk assessments for global enterprises. More than 90% of those organizations had AI applications, models or other servers exposed to the internet. More than a third carried known exploitable vulnerabilities.

So the company built a control plane for machines. AI Broker handles MCP and A2A traffic — how agents call tools and each other — and an Agent Registry tracks what each agent may do. AI Endpoint Security and AI Access Graph round out the stack.

Adoption is early but measurable. Zscaler processed millions of MCP requests a month in its most recent disclosed quarter, up from almost nothing a couple of quarters earlier. Its customer base now holds more than 3,400 AI applications, quadruple the count a year ago.

The business underneath is already large. Annual recurring revenue reached $3,771 million in fiscal 2026, up 25%, and free cash flow hit $779.1 million at a 23% margin. Security-for-AI bookings grew more than 50% sequentially in the fourth quarter, though management publishes no dollar figure for them.

Roughly 70% of those AI-security deals also included Data Security, which is the cross-sell engine working. Metered pricing produced about 30% of new and upsell contract value for the year, and the recurring revenue tied to those offerings more than doubled.

Now the rotation. Zscaler shares jumped about 10% intraday on Sept. 14 as Anthropic's call to pace the frontier pushed chip names lower, with SK Hynix, Samsung, ASML and Infineon each down several percent. Okta, CrowdStrike and Palo Alto Networks moved higher.

The logic is uncomfortably simple. Slowing how fast the industry builds new agents does not slow how many agents already exist, how many identities they hold or how much access they carry. Guardrails get more urgent, not less.

Zscaler still trades well below its 52-week high, and its fiscal 2027 outlook implies growth stepping to roughly 17%. Management also plans to cut about 3% of staff in a restructuring and keeps capital spending elevated on higher memory prices.

The bet is that Zscaler becomes the toll booth of the AI era. Every user, workload and agent that wants to reach an application passes through a checkpoint the company operates. If agents really are the next attack surface, that is a very good place to stand.

Disclosure: The Signal holds no position in ZS. Positions may change. This is not financial advice.