SentinelOne's Weingarten: AI Hasn't Made Attacks Smarter, Just Faster — And Transformers May Be the Wrong Architecture to Fix It
Goldman Sachs Communacopia + Technology Conference, September 10, 2026
SentinelOne founder and CEO Tomer Weingarten used his appearance at the Goldman Sachs Communacopia + Technology Conference to deliver an unusually blunt assessment of the AI security narrative gripping the industry, arguing that recent high-profile incidents reflect speed rather than sophistication, and that the entire transformer-based AI stack may need to be rearchitected before enterprises can truly trust autonomous agents. Alongside CFO Sonalee Parekh, who is in her second full quarter in the seat, Weingarten also detailed a business posting accelerating net retention and record growth in AI security products, following what Parekh called one of the largest guidance raises the company has delivered in recent memory.
No New Techniques, Just New Speed
Asked how sophisticated customers are reacting to headline-grabbing agentic breaches, Weingarten pushed back hard on the premise that AI has introduced novel attack methods. "Fundamentally, there's nothing new," he said. "What these agents have done is nothing that a human attacker has not done in the past. We're not seeing any type of new behavior. We're seeing a new level of velocity, a new level of speed, but we're not seeing new techniques." He was equally dismissive of vendors claiming to have solved the problem: "No matter what you heard, there's nothing that solves the issue we're facing right now in terms of speed and velocity... This notion that we can stop it, we can prevent it, we can live in a world where these things are not happening — it's misguided. We can reduce the risk." His broader point was that AI is exposing decades of poor cyber hygiene rather than outsmarting defenders. "AI is not good at cybersecurity at all. We are bad at building secure infrastructure, all of us combined. AI, which is a great search engine, is able to find all these things that we have basically sucked up for many, many years."
He offered a concrete example from earlier this year involving poisoned open-source libraries such as LiteLLM and Axios, which were unwittingly pulled in by AI coding tools like Claude and Codex, compromising hundreds of enterprises in an automated fashion. SentinelOne's decade-old behavioral detection engine, he said, caught the activity without any prior knowledge of the specific exploit because it was built to flag anomalous behavior generically rather than match known signatures. "The system just saw something. It didn't care if it was Claude or the user or an attacker."
The Case Against Transformers
Perhaps the most striking part of the session was Weingarten's skepticism about the durability of the transformer architecture itself. "I don't think there's any certainty that this is going to be the dominant architecture for years to come," he said, suggesting that AI-assisted coding velocity could enable an entirely new reasoning architecture rather than one reliant on "brute forcing chain of thought and randomizing tokens." He questioned whether the industry's return on ballooning AI compute spend has materialized: "Where are the AI outcomes? Where are they proven in the market? Revenues, dramatic growth, amazing outcomes, transformative change for earth" — rhetorically implying they are largely absent.
Weingarten tied this back to a hardware-level argument, contending that security enclaves running at the same privilege layer as the model they're meant to govern are fundamentally compromised. He referenced the industry's post-CrowdStrike-outage push to move security out of the kernel as a promise that has gone nowhere: "Everybody is going to get out. Nobody is out. What are we, like, 2 years, 3 years from that event. Nobody is getting out." His conclusion was that true AI alignment may eventually require new chip designs or more robust security enclaves that separate compute for security functions from the model itself, noting that even Apple's enclave technology has been compromised.
AI Alignment as the Next Cybersecurity Frontier
Weingarten identified AI alignment — ensuring models behave in accordance with human intent — as the single most important investment theme in cybersecurity for the next five to ten years, calling it a problem the industry has never had to solve before. "The AI alignment issue is going to supersede pretty much every other problem we see in cybersecurity," he said. His prescription is notably not more capable models: "You don't need superintelligence to keep AI aligned. You need something that is more balanced, something that is designed to govern, not to be smart, just to know when intent differs from the exhibited behavior." He was skeptical that guardrails as currently implemented can hold: "The model does what it believes is serving the goal it was given. It looks at the guardrails and says, okay, I see the guardrails, but maybe I'll do this. It's almost like a kid."
Calling Out the Fine-Tuning Hype
Pressed on whether SentinelOne would apply its proprietary telemetry to fine-tune a large language model for offense-defense automation, Weingarten was dismissive of the "red agent, blue agent" narrative popular among vendors and downplayed the value of post-training customization generally. "Doing stuff post-training is very, very limited," he said. "You're not changing the true innate behavior of the model." He singled out Nvidia's Nemotron model by name as an example of the flawed logic behind such efforts: "That thing is not even top 10 in reasoning performance by any benchmark, honestly... to just pick one model and say I'm going to train this to be the best thing of all models, I think that's total wishful thinking." He also made a pointed claim about the competitive standing of Chinese open-source models, saying "I've seen them add incredibly sophisticated ways to scale their models" and that some are "probably like a factor of 5 better" than certain Western alternatives referenced in the discussion.
Fundamentals Over Fairytales
Asked how SentinelOne competes against larger rivals with bigger marketing budgets, Weingarten was characteristically direct, framing the debate as one of substance versus narrative rather than scale. "We show up every day to POCs with the world's leading companies, and we win time and time again, regardless of the size of Microsoft or Palo Alto Networks or CrowdStrike," he said, adding that competitors have "bigger microphones" but not necessarily better technology. He took aim at the platform-consolidation narratives common on rival earnings calls, particularly around identity security and serial acquisitions: "Why do people continue believing in those stories? These stories change. If you just go back 3 to 6 months, it no longer holds... it feels like a cybersecurity theater that's completely disconnected from what's actually happening in the environment."
The Numbers Behind the Message
On the financial side, Parekh reiterated that SentinelOne beat and raised its full-year revenue guidance last quarter, with the analyst on stage noting it was one of the largest beats the company has posted in several quarters; Parekh said the guidance raise exceeded the magnitude of the beat itself, reflecting management's confidence in forward demand. She pointed to net retention strength in the $100,000-plus customer cohort as validation of the platform strategy, along with tripled year-over-year ARR growth in the company's AI security products, Prompt and Purple AI. Parekh also flagged early productivity gains from automating the long-tail renewals process and modest but measurable contraction in sales cycles, both of which she said are beginning to show up in reported numbers. Looking ahead to fiscal 2026 planning with Weingarten, she said a central priority is further reducing churn and downgrades as customers adopt more of the platform, which she argued is the more economically efficient growth lever than new customer acquisition.