2025
2026
Featured in Motions
& Tailwinds
”Complexity is the enemy. The reality is most cybersecurity failures today are not due to lack of tools, they're due to a lot of fragmentation and lack of integration.”
”Identities are increasingly becoming a core control point - whether in cloud, SaaS, or AI. Everything comes down to who it is and what access it has.”
”We’re seeing a revolution happening in the Security Operations Center (SOC) today. There is growing excitement around the ability of agentic AI to automate many manual tasks”
”We're already seeing close to 20,000 agents within a single enterprise.”
”A big prediction for 2026 is that the governance model for AI agents will have to change”
”People are starting to experience AI fatigue.”
”Everyone wants to be AI - not just AI, but agentic AI. You hear it everywhere: “we’re agentic-first,” “we’re agentic this”
”It's very important for leaders to distinguish between what’s real and what’s just narrative.”
”We’ve actually heard from several CISOs that tier-one triage and review have mostly been automated.”
”If your enterprise’s tier-one cases aren’t already automated, you might be behind.”
”We’re already seeing 20–30% of tier-two tasks being automated by agents.”
”For now, you’re not going to see a fully autonomous SOC - we still need human context.”
”Today, many enterprises are deploying AI faster than they are deploying security solutions.”
”I think the risk today is that enterprises could face fully agentic cyberattacks without having solutions that can move as quickly to defend against them.”
Motions &
Tailwinds
Top
Motions
A strong market shift demanding fewer tools with broader platform capabilities, driving vendors to focus on measurable cost-efficiency like "Cost Per Investigated Alert."
Vendors are actively making business, behavioral, and operational context the primary engine for security tools, underpinning decisions like alert triage and privilege grants.
Top
Tailwinds
The accelerated operationalization of AI by adversaries, which forces defenders to adopt AI-Native solutions and automation cycles to keep up.
The rapid, unmanaged spread of employee use of AI tools, which is becoming a top operational risk by bypassing security and data controls.
Deep market skepticism towards generic AI claims, leading to a strong demand for transparency, measurable KPIs, and auditability in all AI-driven products.
The focus has moved from securing AI models to governing autonomous agents operating across enterprise systems. Security now centers on behavior, access, and execution.
Disrupted Categories
Non-human identities are rapidly outnumbering human users across cloud and SaaS environments. This expanded the attack surface and accelerated the shift toward just-in-time access and zero-standing privilege models. Now vendors are beginning to build identity and access control layers specifically for autonomous agents.
A new class of platforms integrating AI and autonomous agents to perform Tier-1 work in Security Operations Centers.
Tools and solutions focused on managing the specific security and safety risks of Large Language Models, including measuring Hallucination Rate and Guardrail Accuracy.
The new architecture of vendor-agnostic security data pipelines, decoupling storage and compute to allow detections to run wherever the data resides.
The AppSec domain is shifting to move critical controls upstream into the developer workflow and secure code that is written or assisted by AI.
2025 -
A Year in Review
AI stopped being an experiment in 2025 and became operational. Across the security stack - in SOC workflows, developer environments, and adversary toolkits - AI moved from pilot projects into everyday use. So, as attackers and defenders suddenly had the ability to iterate faster and test more ideas, traditional security processes became clearly outpaced.
That shift forced real changes in how teams operate. Agentic triage began cutting through alert noise, Federated Detections improved signal across fragmented environments. Just-in-Time Privilege models started shrinking blast radius, and Continuous Red Teaming replaced assumptions with constant validation. At the same time, resilience and recovery stopped being aspirational goals and became measurable expectations for boards and leadership teams.
Agentic triage refers to a security approach in which autonomous or AI-driven agents actively analyze, prioritize, and route risks or alerts without requiring constant human interpretation.
Federated detections are security detections executed across distributed data sources without centralizing the underlying data.
Just-in-time privilege models grant temporary elevated permissions only when needed and revoke them immediately after use.
Continuous red teaming is the ongoing simulation of adversarial attacks used to identify weaknesses and improve the safety and robustness of AI models.
This report relies on three signal types: direct feedback from CISOs, observable vendor roadmap shifts, and the operational lessons of major incidents throughout 2025. It focuses on what truly changed - and what those changes mean for 2026. The report is organized in three parts: the trends that defined 2025, five technology pivots grounded in clear signals and measurable KPIs, and practical recommendations for the year ahead.
This report is based on the accumulated annual research work of the SACR and Deutsch & Co. research teams, which included 200+ CISO interactions and recorded interviews, 100+ founding team interviews, and internal reviews of 50+ teams backed by industry defining VCs.
It is meant to give an overall review of 2025’s motions and trends, to help security professionals, founders and investors navigate the tailwinds of 2026.
Software Analyst Cybersecurity Research (SACR) is an independent research and advisory organization focused on helping CISOs, founders, investors, and security teams understand where cybersecurity is heading. Through in-depth industry reports, analyst research, vendor analysis, and shorter thought pieces, SACR analyzes emerging technologies, market shifts, and vendor strategies across key security domains. The firm was founded by Francis Odum, Founder, CEO and Chief Cybersecurity Analyst, who built SACR into one of the largest independent cybersecurity research platforms in the market. He is recognized for his work with over 60,000 security professionals worldwide and for establishing SACR as a trusted brand among CISOs and leading vendors.
Deutsch & Co. is a private equity firm focused on investing in and building category leaders across cybersecurity and AI. The firm’s investment strategy is grounded in proprietary research, including hundreds of annual interviews with industry leaders, buyers, and practitioners, used to identify emerging market gaps and define new categories. By combining research-driven insights with strategy, positioning, and branding, Deutsch & Co. partners with companies to help shape and lead the categories they operate in. The firm was founded by Roei Deutsch, who serves as CEO.
Overall Trends:
What Dominated
2025
Throughout 2025, recurring themes emerged from private CISO forums, executive briefings, and peer discussions, revealing both the risks that dominated security leaders’ attention and the priorities shaping their outlook for 2026. Here are the highlights.
AI warfare
context
In 2025, the industry conversation shifted from “security for AI/LLMs” to securing agentic platforms themselves. Early discussions focused on model risks such as prompt injection, hallucinations, and data leakage. But as enterprises moved beyond experimentation and began embedding AI into real workflows, the conversation rapidly shifted. LLMs stopped being static assistants or chatbots and began evolving into agents, capable of taking actions across systems, chaining tools together, and operating with increasing autonomy.
This shift reframed the security challenge: the problem was no longer just securing models, but governing the behavior, access, and execution of autonomous systems interacting with enterprise infrastructure. By the end of 2025, it became clear that the next phase of cybersecurity will center not simply on AI safety, but on securing agentic systems operating inside real enterprise environments.
As organizations began operationalizing agentic systems inside real enterprise environments, a deeper challenge quickly emerged: governing the identities and actions of machines operating at scale. Traditional IAM and SSO architectures, designed for predictable human logins, struggle to manage agents that execute high-frequency, non-deterministic tasks using long-lived credentials across multiple systems. As a result, 2026 will see the emergence of Agentic Identity Access Platforms (AIAP): a new identity control layer that acts as an “SSO for Agents,” brokering task-scoped, ephemeral identities based on agent intent. In the agentic era, identity security will shift from verifying who is acting to continuously governing why an action is occurring and how long access should exist.
In the agentic era, identity security will shift from verifying who is acting to continuously governing why an action is occurring and how long access should exist.
Company Launches
Amid the excitement of embracing the next wave of generative AI, companies are moving quickly to deploy agentic use cases, often overlooking the critical need to secure these systems and control the sprawl of non-human identities.
AI agents are quickly becoming a new class of workforce in the enterprise, but they require more complex identity lifecycle management than human users.
AI agents don't just authenticate, they take action, call APIs, chain workflows, and make decisions. Securing them requires treating identity as a runtime control plane, not a one-time configuration.
Enterprises are moving beyond simple AI chatbots to fully autonomous agents - but with this evolution comes an exponential rise in security and safety risks. The threat vector has escalated from basic prompt injection attacks to mass data exfiltration, supply chain attacks, and even autonomous chaos.
Consistent CISO Concern of 2025
Across CISO dinners, closed-door forums, and year-end briefings, Shadow AI emerged as the most cited operational risk. AI tools proved useful enough that adoption quickly outpaced security’s ability to inventory and govern them. Employees uploaded sensitive data into copilots, browser extensions, and SaaS AI features with little visibility into data flow or retention.
Shadow AI is the unauthorized or unsanctioned use of AI tools within an organization without formal oversight from IT or security teams.
This wasn’t reckless behavior - AI was becoming embedded in everyday workflows, making its use inevitable. Blocking it outright didn’t work; it simply pushed activity into unmonitored channels. The real concern was the absence of control, logging, and policy enforcement. Shadow AI became a symptom of a deeper issue: security teams losing visibility into how work actually happens.
By 2025, the conclusion was unavoidable: AI is here. Security teams must navigate it - without losing control.
Security leaders described a tension between innovation and risk. Development teams moved quickly, leadership encouraged experimentation, and security teams were asked to approve systems they barely understood yet.
The most effective organizations responded by narrowing the scope. They permitted AI in defined domains, such as SOC triage, documentation, and internal tooling, while enforcing guardrails on data access, actions, and auditability.
The lesson for 2026 is not to slow AI, but to effectively constrain the boundaries in which it operates.
AI-Driven Cyber Warfare
2025 was a transition year: security programs began a slow pivot away from non–AI-native stacks, just as adversaries operationalized AI tooling at scale. The resulting shift in attack patterns - faster iteration, lower-cost experimentation, and compressed time-to-breach - made legacy approaches objectively insufficient. For CISOs, this became an additional, concrete driver for modernization: when both sides are automated, advantage accrues to the party that closes the loop first.
In practice, “AI vs. AI” is simply competing automation cycles: faster learning, tighter feedback, and broader execution.
That acceleration also shifted the core risk from data theft to operational paralysis. Incidents like the Jaguar Land Rover attack (estimated $2.5B impact), SaaS-Jacking campaigns targeting major SaaS platforms, and the Salt Typhoon espionage activity showed attackers increasingly optimizing for continuity disruption and deep infrastructure access - not just exfiltration. The implication is blunt: security failures are now business-stopping events, not merely confidentiality breaches.
A 2025 cyberattack that forced Jaguar Land Rover to halt production for several weeks, disrupting global supply chains and causing an estimated $2.5B economic impact.
A Cyberattack in which attackers hijack SaaS accounts or cloud applications through stolen credentials, tokens, or misconfigured integrations, allowing persistent access to organizational systems
The cyber-espionage campaign linked to the Chinese state-backed group “Salt Typhoon”
This reality resulted in two new priorities for security teams: resilience and recovery.
Organizations moved away from periodic assurance toward continuous validation, treating security posture as something you prove, not assume. Continuous red teaming and autonomous attack simulation began shifting from “advanced program” to an operating baseline, continuously exercising real attacker paths and measuring whether controls still prevent, detect, and contain under evolving tactics. Just as importantly, teams started validating execution, not just documentation: playbooks were drilled through repeatable technical exercises to surface latency, ownership gaps, and brittle dependencies - turning response into a practiced capability rather than a binder on a shelf.
In 2025, cyber resilience increasingly became the board’s yardstick for security performance: assume incidents will occur, and measure how quickly the business can restore critical operations. That reframed investment toward verified, attack-resistant recovery - tiered RTO/RPO commitments, dependency-aware restoration sequencing, and routine proof that backups are recoverable and operationally usable. Teams also reduced human latency by operationalizing response and recovery workflows through orchestration and automation, turning playbooks into executable procedures rather than documentation. Mature programs now manage recovery readiness like an SLO: restoration time for Tier-0 services, recovery success rate, and time-to-restore under realistic conditions - continuously validated and extended across critical third- and fourth-party dependencies.
Company Launches
We're moving from the analyst's hands on the steering wheel to autonomous actions that are effective, safe, and reliable. We need full autonomy; it's not optional, it's foundational.
I believe within the next few years virtually all cyberattacks will be AI-based - swarming, tailored, and relentless. They will be untethered to human limitations and capable to execute on a scale we have never witnessed before.
Attackers don't wait for your annual pentest. Neither should your defense. What security teams actually need are high-signal findings they can trust: novel vulnerabilities that are proven exploitable.
Fully autonomous testing tools promise efficiency but introduce security risks and inaccuracies in production environments. Traditional pentesting tools force testers into manual workflows, limiting scalability. Resolve this tension by enabling autonomous pentesting to scale through human-governed AI execution.
Across virtually all sectors, vendors, customers, and domain experts are signaling the shift toward contextual security. Alert triage and enrichment are increasingly executed through the lens of business-critical context; just-in-time privileges are being issued and revoked based on usage context and behavioral pattern recognition; and secure coding is shifting toward context-aware guidance - surfacing prescriptive fixes directly from pull requests and code reviews.
Context is no longer merely an input to security decisions; it now underpins them
Its influence is pervasive today and is poised to become even more dominant over the coming year.
By 2025, security tooling fragmentation was becoming harder to justify: many teams were running broad stacks of overlapping tools across endpoint, identity, cloud, network, and detection. That sprawl added unnecessary costs and operational burden (more integrations, more policy surfaces, more triage), yet left teams with hard-to-trace visibility gaps.
At the same time, we started seeing clearer consolidation signals. Among others, Google’s acquisition of Wiz suggests that leading point solutions are increasingly being treated as “platform-grade” capabilities. Despite ongoing concerns around vendor lock-in, we expect 2026 to extend this direction: more stack rationalization and more vendor consolidation, with a stronger emphasis on fewer tools that drive prioritized remediation over noisy detection.
This trend is amplified by the previous topic - contextualization - as the trend of putting AI & organizational context in the core of each security product, brings the different tools - and traditionally-defined cybersecurity quadrantable categories - even closer together.
Company Launches
By mid-2025, CISOs were openly dismissive of generic AI and LLM claims. Many vendors added “AI” to their messaging without fundamentally changing their products, and hype alone no longer inspired trust. CISOs demanded both proof of the advantages of the LLM through clear KPIs and a measurable way to audit, understand, and control its behavior. Vendors that couldn’t provide precise answers, validation, or evidence quickly lost credibility. Adoption would no longer be driven by marketing claims - security leaders wanted results, not buzzwords.
During 2026, the term “AI agent” or “agentic LLM” is prone to suffer the same fate, as CISOs now see past the AI hype, demanding demonstrable value and explainability.
Distinguished
Tech Pivots
The following five areas capture the main structural changes that defined 2025 and now shape 2026 planning and execution.
The Emergence of AI SOC
A year ago, every meeting started with: "Does AI actually work in security? "Today, it's: "How do I operationalize AI agents in my SOC?"
The old approach of configuring and maintaining endless playbooks doesn't scale. Attackers are already using AI to launch bigger and faster campaigns. Security teams need tools that don't just keep up but actually learn and improve continuously.
This is not about eliminating jobs. It's about ensuring an analyst doesn't have to spend time triaging and investigating alerts, because who wants to do that all day, every day? Instead, they can focus on the 4% of issues that truly matter to an organization.
What we actually want to measure is that the AI is accurate, comprehensive, and that it takes on work that is actually valuable. If it is, measuring how many equivalent analyst hours are done by the AI is a great metric to start with.
Enterprises now operate in a world where anyone who knows how to talk knows how to hack.
AI is here to stay, and enterprises must implement strategies to monitor and protect AI use. Traditional security offerings were not designed for the ways AI applications operate.
Everyone talks about AI, but AI is really only as useful as the tools and the resources it has access to.
Enterprises we’re working with have 50 to 200 LLM applications today. That number could double, and then quadruple, in just the next few years.
The Security Data Layer
Enterprises aren't just overwhelmed by data volume; they're being outpaced by its complexity
Security is, at its heart, a data problem, and legacy, rules-based data pipeline platforms simply weren't built for today's ever-growing attack surface and data-rich security operations
The current operating model of the SIEM - the dominant technology in this domain for the last two decades - is not only 'crazy expensive,' but is also increasingly causing AI-native security operations to fail.
I think we are discussing AI [too much] and losing the context. AI is changing our lives, but perhaps not [yet]. We want to show the market that data is the only place where all tools, all attacks, and everything are together.
I hear fewer questions about “what platform should we choose?” and more about “how do we manage our data so each tool gets what it needs to do its job well?
From Users to NHIs & Agents
Identity is under relentless attack, and adversaries are going straight for the keys to the kingdom — privileged access. From social engineering to sophisticated insider abuse, they're escalating privileges to access the most sensitive systems and data.
We're at a pivotal moment in identity security. The unseen dark matter of identity is overtaking what organizations can manage or even see. It's no longer about control - it's about context.
For years, companies assumed the root of identity security was making access as convenient as possible. But what has changed is the scale and dynamism of modern environments. Humans can manage things manually, but organizations operating at today's speed, especially with AI agents, need systems that can handle constant change.
In a world where AI is transforming software development, the biggest security risk isn't just in the code - it's in how the code is written.
The problem is clear: AI has pushed engineering velocity far beyond what reactive security tools were built to handle.
Making security a positive experience for developers is key to growing their cyber judgement and knowledge. By integrating AI-powered training into their workflow and using their current work as the reference point, developers learn in a way that's impactful, helping them better understand and resolve security vulnerabilities without disrupting productivity.
Newly emerging
tailwinds
While there were no meaningful public motions to reference in these categories, having seen many teams working in stealth on solutions to these and given the strong technological shift that enables/requires them - we make an educated guess that these categories would be prominent in next year’s report:
Cross Cutting
Recommendations
For CISOs
These recommendations reflect lessons from 2025 as AI adoption accelerated across security and engineering environments. As AI capabilities expand, organizations must adopt them in ways that remain measurable, controllable, and auditable. Risk is reduced when clear operational boundaries are defined, when the full lifecycle from development through production is hardened, and when real incidents are systematically converted into stronger preventive controls. Sustained executive support depends on translating resilience into business-relevant outcomes that leadership can track and fund.
Start with a narrow, high-frequency security use case where outcomes are measurable and the blast radius is controlled (e.g., phishing triage, alert summarization, investigation drafts). Running AI in a contained domain allows teams to validate reliability, measure productivity gains, and understand operational risks before expanding its role. Once the system consistently improves triage speed, investigation quality, or analyst workload, the scope can gradually expand to additional workflows with higher operational impact.
Require guardrails across the entire AI lifecycle, from development to production, including strict data policies, versioned prompts and tools, approval workflows for sensitive actions, and full logging of AI decisions and evidence. AI systems operating in security environments must function under controlled autonomy, where their behavior can be audited, explained, and constrained.
Don’t treat telemetry ingestion as a free resource. Uncontrolled data pipelines create noise and drive up cost without improving detection quality. Instead, treat each telemetry source as a measurable investment: track the ingestion, storage, and compute costs alongside the detections and investigations it enables.
Ensure that identity risk is visible inside the operational environments where access decisions are actually made, not only within security dashboards. Instead of relying solely on centralized IAM or governance tools, make sure there is an integration of identity risk signals into developer and platform workflows. This allows engineers and platform teams to see the security implications of privileges, tokens, and access paths at the moment they create or modify them.
Don’t assume API tokens, service accounts, and machine identities are low-risk or temporary. In modern cloud and SaaS environments, these identities often hold persistent privileges and can create powerful attack paths if left unmanaged. CISOs should ensure that non-human identities are governed with the same rigor as human credentials, including automated discovery, rotation policies, and lifecycle controls across the SaaS and API ecosystem.
Translate cyber resilience into clear, measurable outcomes that leadership and the board can understand and track over time. While no universally accepted KPI exists for resilience, CISOs should prioritize indicators that demonstrate tangible improvements in operational performance and risk reduction. The specific metric matters less than its transparency, repeatability, and connection to business impact. Frame progress through trend lines and measurable changes in exposure, rather than relying solely on abstract technical metrics.
Employees will adopt AI tools whether security teams approve them or not. Instead of attempting to block usage entirely, CISOs should focus on discovering, monitoring, and governing AI adoption before it evolves into unmanaged risk.
Security improvements did not come from adding more tools this year, they came from making them work together.
The programs that improved fastest were those that connected signals, shared context across controls, and enabled teams to act quickly on what they saw.
AI did not change that principle. It accelerated everything - attacker experimentation, alert volume, and response timelines - but the difference between strong programs and weak ones remained coordination. Systems that share context and support decisive action outperform stacks that simply accumulate tools and signals.
With the goal being to reduce time between risk appearing and controls taking effect, faster detection and containment, quicker restoration of services, and a smaller exposure window for revenue-critical systems became both the method, and the metric.
