SentinelOne Says AI-Assisted Hunting Now Surfaces 28 Customer Intrusions a Day, Up 2,200% in a Year, as It Cites Frontier-Model Cyber Capability Doubling Every 4 Months
OneCon 2026 pre-show session, Las Vegas, October 20, 2026
SentinelOne used the opening evening of its OneCon customer conference to put numbers behind its AI-assisted threat hunting program, which launched at last year's event. Chief Customer Officer Steven Stone presented the figures as an accountability exercise, a look back at what the company promised on stage a year ago. The main product and strategy announcements are scheduled for October 21, with product leadership and CEO Tomer Weingarten set to present. Tonight's session was narrower, and the disclosures were operational rather than financial. They still give investors a rare quantitative view of how SentinelOne's newest AI capability is performing in production.
AI-assisted hunting scales fast, though from a small base
The central data point is that AI-assisted hunting, which went live on December 1, 2025, now notifies customers of intrusion activity 28 times a day. Stone said the company's ability to find malicious activity with the method rose 380% in its first quarter, 1,400% at six months, and 2,200% over the full year. He was candid that the early figures carry limited weight. "Any of us that have built something, your first iteration, you're going from nothing to something," he said. "That's not really a meaningful number." The 28 daily notifications is the figure he said matters to customers.
Stone also said SentinelOne identified more than 14 million true positive detections for customers in the last calendar year, a 50% increase. He set that against roughly 16% growth in endpoints per organization, which he said matches SentinelOne's own installed base. Detection volume is therefore growing about three times faster than the surface area being protected. He attributed this to the company's autonomous models. He noted that SentinelOne's first patent, filed in 2014, covered an autonomous malware detection model, and he argued that this history gives the company a head start while others struggle to move into AI. The company provided no revenue, margin or attach-rate data tied to these capabilities, so how the technical progress translates into financial results remains open.
React2Shell: a stress test within 48 hours of launch, with an admitted Windows gap
Stone described the React2Shell vulnerability as the first real test of the new hunting engine. The flaw, disclosed on December 3 with a CVSS score of 10.0, allowed compromise through a single crafted HTTP request with no authentication or user action. SentinelOne launched AI-assisted hunting two days before disclosure. Stone said the company observed the first React2Shell activity at a customer on December 9, and that AI-assisted hunting was the first mechanism to flag it. The hunting engine had been built to recognize the general shape of web exploitation.
Stone was unusually direct about weaknesses. "We knew we have strong coverage in Linux and Mac early, but we had real challenges on Windows," he said, adding that the company had "weaknesses we had to cover." He also made a point that investors weighing the AI-versus-human debate should take seriously. The first alert looked like routine penetration testing, and it was a human hunter with years of pentest experience who judged that something was wrong. "We love to say we have technology answers, everything is universal. It's machine driven. It's not how we got going on this," he said. "We had an expert who looked at pentest for years, and that expert could say, I just know this isn't right."
Stone explained why this type of detection is deployed as hunting rather than as a production detection. A tool that recognizes the broad shape of web exploitation would, in his words, "explode" on endpoints and "bury our customers in alerts." Findings from hunting are refined and then converted into production-grade detections, which in turn frees the hunting layer to pursue new activity. Over the first 30 days, SentinelOne identified 8 attempted victims among its customers, across activity that included domain reconnaissance, which Stone said looked designed to lead to deeper intrusions, and cryptocurrency mining. SentinelOne's own research identified at least 7 significant threat groups exploiting the flaw in the first week. Stone cited AWS reporting a China-based group within about 12 hours, and noted that CISA's addition of the flaw to its Known Exploited Vulnerabilities list came 36 hours after first exploitation.
SectopRAT shows the less glamorous use case
The third example was deliberately mundane. SectopRAT, a browser session and crypto wallet stealer that SentinelOne has tracked for 7 years, ships in a new delivery wrapper every few months. The most recent wrapper masquerades as a fake Claude Desktop installer. Stone said AI-assisted hunting was not the core detection method here. It served as a quality-control layer, checking that static rules, behavioral detections and the roughly 50 known command-and-control indicators were still catching variants. The company also built a DLL side-loading hunting pack aimed at the behavior rather than the specific malware. This suggests the near-term value of the technology is as much about maintaining coverage as about discovering novel threats.
Threat backdrop: offensive AI capability tripling, attackers at machine speed
Stone cited a UK government assessment, published about three months ago, that frontier models are doubling their offensive cyber capability every 4 months. By his arithmetic, that capability has tripled since the previous OneCon. He said attackers' use of AI is no longer a prediction. "That's not next year, that's last year," he said, noting use by individuals, small groups and governments.
He pushed back on the view that AI-driven vulnerability discovery is the main problem. "Finding vulnerabilities was never the attacker bottleneck," he said, estimating that defenders were "multiple" vulnerabilities and exploits behind 20 years ago, 100 behind two years ago and 1,000 behind now. The bigger shift, in his view, is that "for the first time, we are literally seeing attackers move at machine speed." He offered a measured upside: AI-driven attacks may be easier to detect because agents are less disciplined about evasion, and prompts force attackers to "literally write out what they want to do."
Defender-advantage claim and what it implies for the platform
Stone ventured that this is "the first technology I actually think benefits defenders more than it does attackers," while stressing that tools alone are insufficient without experts. That framing underpins SentinelOne's pitch around its relaunched threat services, the Wayfinder offering, Purple AI and agentic investigations. Stone said agents can be assigned tasks like staff, running around the clock and lifting analysts out of routine work. The claim is plausible and consistent with the company's metrics, but it remains the view of an executive at a vendor selling the solution. Notably, the session offered no independent benchmarks, no false-positive rates for the hunting engine, and no customer-level economics. Investors will look to tomorrow's product roadmap and Weingarten's remarks for evidence that these operational gains can drive durable growth in the platform business.