SentinelOne

2013: why traditional antivirus stopped working

SentinelOne was founded in 2013 by Tomer Weingarten, Almog Cohen and Ehud Shamir, and Weingarten has served as CEO and chairman ever since. The idea it was built on is the opposite of classic antivirus. Traditional antivirus software works by signatures: it keeps a list of digital fingerprints of known malware and compares every file that reaches the computer against it. This method worked well as long as the number of malware programs was limited and creating a new one took effort. It collapsed when it turned out that a small change to the code of existing malware (minutes of work, and later fully automated work) was enough to produce a new signature that appears on no list. In other words, the defense is always one step behind by definition, and attackers learned to take that step cheaply. This also has a simple structural consequence: when the number of variations that can be produced automatically is far larger than the pace at which the list is updated, the method breaks. Not because the idea is wrong, but because it does not scale, and it will always be chasing whoever has already attacked.

An agent that watches behavior and decides on its own

Instead of comparing files to a list, SentinelOne runs an agent on the computer itself that analyzes in real time what the software is doing: which files it opens, where it connects on the network, which new processes it creates, whether it starts encrypting files one after another. The decision is based on the pattern of behavior, not on the identity of the file. So malware that was written five minutes ago and that no one has ever seen still looks suspicious when it behaves like malware. The second part, and the less obvious one, is that the agent is authorized to act on its own: to disconnect an infected computer from the network or to restore encrypted files to their previous state, without waiting for approval from a human analyst. In a ransomware attack, the difference between an immediate response and a response twenty minutes later is the difference between one infected workstation and an entire department. The same technology is also applied to internet of things devices and to cloud workloads. This approach also has a downside, and it comes from the same feature: a system that decides on its own can also be wrong on its own, and disconnect a computer that did something legitimate that looked suspicious. The balance between detection sensitivity and the number of false alerts and actions is one of the key settings security managers tune in products like this.

An important clarification: the AI here is not a language model

When a cybersecurity company says artificial intelligence, it is easy to assume it means something like ChatGPT. That is not the case here, and the difference is fundamental. What runs in SentinelOne's agent is statistical models for classifying behavior: systems trained on huge numbers of examples of legitimate and malicious behavior, which learn to tell them apart. There is no conversation here, no text, and no attempt to understand human language. This is a much older branch of machine learning, less glamorous than the generative models that star in the headlines, but it runs quietly on hundreds of millions of devices and represents the most mature and commercial part of the market. This distinction is worth remembering every time a product is advertised as AI-powered. It is true in many cases, but it describes a whole family of technologies that are very different from one another. The distinction also explains why the company needed a separate acquisition to enter the field of language model security, as described below. It simply was not the same kind of knowledge, or the same people. Among cybersecurity companies that do not rely on language models there are also different approaches: the Israeli company Deep Instinct, for example, presented itself as the first company to apply end-to-end deep learning to cybersecurity, with the aim of stopping attacks even before they run.

From IPO to a string of acquisitions

The company's funding came quickly: $120 million in a Series D in June 2019 led by Insight Partners, $200 million in a Series E in February 2020 at a valuation of $1.1 billion, and later that year another round of $267 million that raised the valuation to $3.1 billion. On 30 June 2021 the company went public on the New York Stock Exchange under the ticker S and raised $1.2 billion, at the time the largest cybersecurity IPO ever. Before and after the IPO it made a string of acquisitions that widened the platform from endpoint protection alone into a broad security system: Scalyr in February 2021 for $155 million, Attivo Networks in March 2022 for $616.5 million, the consulting group Krebs Stamos in November 2023 for an undisclosed amount, and PingSafe in January 2024 for more than $100 million. In September 2025 Observo AI was added too, for an undisclosed amount. The overall pattern is clear: the company bought data collection, cloud protection and consulting capabilities, and assembled them into a broad platform around its core product.

Prompt Security: when a cybersecurity company buys AI know-how instead of building it

In August 2025 SentinelOne bought the Israeli company Prompt Security for about $250 million, its entry into the field of generative AI security. This field is fundamentally different from what the company had done until then: the risk is not a malicious file that reaches the computer, but a prompt an employee sends to a language model that exposes sensitive business information, or an answer the model returns that contains something that was not supposed to get out. There is no file to scan and no process to isolate; there is text flowing between an employee and a model, and it has to be monitored without paralyzing its use. The move reflects a broad pattern that has repeated over the past two years: established cybersecurity companies, which have customers and distribution channels, buy AI know-how instead of building it from scratch, mainly because the field is moving too fast to catch up from the inside. This acquisition also adds a layer to the Israeli story: a company founded in Israel, trading in New York, buying another young Israeli company, a cycle that is becoming more and more common in the local cybersecurity industry.

A billion dollars in revenue, and still a loss on the bottom line

In the fiscal year that ended in January 2026, the company crossed $1 billion in revenue for the first time: $1,001.3 million, up 22%. Annual recurring revenue stood at $1,119.1 million, also up 22%, and the number of customers paying more than $100,000 a year rose 18% to 1,667. Alongside these numbers, it is important to read the bottom line too. In that same year it recorded a net loss under generally accepted accounting principles (GAAP) of $450.7 million, and its GAAP operating margin was negative, minus 32%. The company's announcement of operating profitability refers to an adjusted, non-GAAP measure. This is typical of growing software companies, yet the measures should be told apart: the company is growing fast and profitable by one measure, and losing a considerable sum by the binding measure. The explanation for this is accepted in the industry: companies at this stage prefer to invest every extra dollar in sales and development to grow fast, and see profitability as a goal for a later stage. The question that remains open for investors is when exactly that stage arrives.