a16z Survey of 100 CISOs: AI Reshapes Enterprise Security Agenda, Nearly 90% of Budgets Expected to Grow, Six Major Offense-Defense Domains Enter a Reshuffling Period

Authors: Steph Zhang, Owen Colby (a16z)

Compiled by Shenchao TechFlow

Shenchao TechFlow Editor's Note:For a long time, the cybersecurity industry has been preparing for the possibility that quantum computing could upend asymmetric encryption. Yet no one anticipated that large language models and autonomous AI agents would surge forward at a pace beyond Moore's Law, first detonating shockwaves in software engineering and penetration testing. From OpenAI agents unexpectedly escaping during security evaluations and penetrating Hugging Face production systems, to automated vulnerability discovery lowering the barrier to zero-day exploitation to a mass-market level, AI is arming attackers and defenders alike with extremely asymmetric efficiency.

On the eve of this dramatic shift in offense and defense paradigms, top Silicon Valley venture capital firm a16z conducted an in-depth survey of 100 chief information security officers (CISOs) at leading companies with more than 500 employees. The survey found that 87% of security leaders expect net growth in security budgets for fiscal year 2027. The core focus of enterprise defense is undergoing a historic migration: from passively accumulating alerts and stopping at detection, to fully automated closed-loop remediation through patching; from quarterly manual penetration testing, to continuous 7x24 autonomous red-team offense and defense.

At the same time, internal shadow AI and the abuse of privileges by AI agents have created unprecedented new attack surfaces, while the widespread adoption of AI-assisted coding tools has caused codebases to expand exponentially, sharply raising the bar for static application security testing (SAST) and runtime protection. This report provides a comprehensive view of the actual procurement intentions and budget allocation priorities of 100 CISOs across core dimensions such as offense and defense, SOC operations, data security, and platform consolidation, offering a forward-looking industry compass for cybersecurity entrepreneurs, enterprise technology decision-makers, and primary-market investors.

a16z General Partner Martin Casado published a new long-form article today, deeply analyzing the four foundational philosophies that guide investors in evaluating cybersecurity procurement and build.Click here to read the full article。

For years, the entire cybersecurity industry has been bracing for the day when quantum computing would completely overturn the rules of network defense. However, artificial intelligence (AI) knocked on the door first.

This summer, OpenAI agents unexpectedly escaped during a cybersecurity benchmark evaluation and successfully penetrated Hugging Face's online production systems. The agents were supposed to reason about vulnerabilities in a sandboxed environment, but they traced a path to valid credentials and crossed isolation boundaries to take control of critical infrastructure.

And this was only the tip of the iceberg in a much larger unfolding. Since then, AI has begun, at an unprecedented pace, to continuously help hackers and security researchers uncover deeply buried weaknesses in the foundational software of the global internet.

Just last month, a research team used frontier large models to comprehensively audit open-source codebases and uncovered more than 20 previously unknown zero-day (Zero-Day) vulnerabilities. Some of these were even in legacy code that had been running smoothly in critical production environments for more than 20 years.

Code defects have always objectively existed, but the core qualitative change now is that large models are becoming extraordinarily good, with exponential efficiency gains, at finding weaknesses in massive codebases and autonomously deriving viable exploit chains.

Figure: AI-assisted vulnerability discovery efficiency shows a generational leap

At the same time, the number of disclosed high-severity vulnerabilities has also surged. Facing a flood of alerts, already stretched-thin enterprise security teams are seeing the time they can allocate to each incident response compressed to the limit.

Figure: Security vulnerabilities and high-severity vulnerabilities facing enterprises have surged year after year

More troubling is that the digital perimeter security teams are responsible for protecting is itself expanding rapidly. With various AI-assisted tools, ordinary employees are developing and launching new applications at an unprecedented pace; business systems that once took specialized architects months to build can now go live in weeks or even days.

Enterprise data is being accessed repeatedly at unprecedented scale and frequency. Departments are feeding proprietary documents into large models and building complex automated integration flows between internal SaaS tools and business data.

This helps explain why, in secondary markets, the total market capitalizations of leading cybersecurity platform giants such as Palo Alto Networks and CrowdStrike continue to reach record highs, and why next-generation cybersecurity unicorns are completing mega funding rounds at ten-billion-dollar valuations at a startling pace.

Chart: Total market capitalization of leading cybersecurity platforms has continued to surge in recent years, setting new all-time highs

The boundaries of assets requiring protection are expanding in an uncontrolled manner, while large enterprises are swiftly and unhesitatingly increasing their security budgets.

To understand how frontline security decision-makers are responding to these unprecedented pressures, we conducted an in-depth quantitative and qualitative survey of 100 chief information security officers (CISOs) at large enterprises with more than 500 employees. Responding companies spanned mainstream pillar industries including financial services, healthcare, advanced technology, and high-end manufacturing.

Unsurprisingly, the vast majority (87%) of CISOs clearly expect cybersecurity spending to continue growing strongly in 2027. The core engine driving this budget expansion is the escalating pressure from AI-enabled external threats, as well as the non-negotiable need to protect enterprises' increasingly proliferating in-house AI applications.

Chart: Survey of 100 CISOs: 87% of respondents expect total cybersecurity budgets to continue growing

Among these new capital flows, the majority are explicitly allocated to purchasing modern security tools. A particularly encouraging signal is that the surge in tool budgets has not come at the expense of staffing.

Nearly half (48%) of surveyed enterprises said they plan to expand software procurement while simultaneously hiring more full-time security specialists and analysts.

Chart: Breakdown of new security budget allocation: tool procurement leads while hiring continues to expand in tandem

What is the core driver behind this simultaneous growth in both technology and headcount? CISOs' true ambition is to comprehensively expand the coverage of their defense perimeter: investigating edge alerts that previously had to be deprioritized, conducting exhaustive security testing of rapidly multiplying in-house applications, and significantly shortening the window between vulnerability exposure and full remediation.

CISOs at top U.S. enterprises are no longer merely making plans on paper. They are actively running product proof-of-concept (PoC) and pilot tests, earmarking dedicated procurement funds, and carefully deciding which products they trust enough to deploy into live production environments.

Many major procurement decisions will be finalized over the next two to three quarters.

AI Is Reshaping Enterprise Security Priorities

In today's cybersecurity world, a single year is enough to produce generational shifts.

When we asked security leaders to review their early 2026 defense priorities, Exposure & Vulnerability Management, Cloud Security Posture Management (CSPM), and Application Security (AppSec) firmly topped the priority list.

Yet just months later, when they redrew their 2027 blueprints, the priority landscape shifted dramatically. Continuous offensive security testing (Offensive Security), security governance for internal AI agents, and automated Security Operations Centers (SOCs) surged to become the industry's most prominent focus areas.

Which initiatives are classified as the most urgent P0-level strategic priorities depends on each company's specific architecture. Automated SOC operations centers received the most first-place votes; meanwhile, other teams ranked offensive security or internal AI agent controls as their top priority.

Chart: Survey of 100 CISOs: Distribution of Core P0 and P1 Enterprise Security Priorities for 2027

The shift from the established 2026 priorities to the new 2027 blueprints vividly illustrates the remarkable speed at which the entire security agenda is being disrupted and rebuilt.

Among them, internal AI agent security, automated SOC response, and continuous offensive engagements emerged as the three key breakout categories with the largest ranking gains across the industry.

Chart: The core categories gaining the fastest CISO attention: AI agent security, SOC automation, and continuous offensive engagements

This historic shift in priorities is fundamentally driven by two deep-seated imperatives that cannot be ignored:

First, enterprises must decisively secure the boundaries around AI tools already widely used by employees and customers.

Second, defense teams must use AI to counter AI, fully integrating AI productivity into their daily defenses to keep pace with the rapid evolution of attackers' arsenals.

The next procurement window has opened.

While market debate over frontier AI security capabilities has reached a fever pitch, the online defense systems of most traditional enterprises remain in a transitional pilot-and-validation phase. Only 18% of surveyed companies reported that their core defense pipelines have fully deployed production-grade AI-native capabilities.

However, the inflection point is approaching rapidly: most companies have already launched in-depth pilot tests and clearly expect to move them into formal production environments within the next year. Among them, code security review and intelligent email defense currently lead in production-level adoption rates.

Figure: Adoption stage distribution for key AI security technologies: pilot testing is accelerating toward full production-grade deployment

Compared with lengthy system replacements, enterprises’ actual procurement plans often move much faster.

Survey data shows that as many as 45% of CISOs explicitly plan to complete new purchases or major upgrades for internal AI agent security within the next 12 months; more than one-third of decision-makers are also preparing to sign substantial procurement orders for offensive security and automated SOC tools.

Figure: Technology categories CISOs plan to newly procure or significantly upgrade within the next 12 months

However, what is highly challenging is that buyers’ decision processes are becoming intensely intertwined with the rapid iteration of the products themselves. Some buyers are expanding first-phase pilots, while more buyers are struggling to weigh options amid dazzling new-feature demos, trying to lock in the most promising path before the technology converges.

Below is a full picture of the six industry-evolution trends revealed to us by 100 top CISOs, along with the disruptive startup opportunities we firmly believe in.

Offensive security is evolving into a continuous, always-on capability.

Earlier this month, a16z led the Series B round of Armadin, founded by prominent security leader Kevin Mandia. We are making a major bet on this category because we firmly believe defense teams must think like the best attackers and possess equally agile offensive operational capability.

On a rapidly changing cyber battlefield, speed is a matter of survival. Under traditional penetration testing models, security teams often spent weeks reviewing report findings item by item, then waited months or even longer before scheduling the next comprehensive retest. But in fast-changing multi-cloud architectures, such low-frequency static reviews are a futile exercise.

The explosive evolution of AI has multiplied the lethality of this time gap. A recent estimate by consulting giant Bain indicates that AI tools can compress the typical attacker kill chain—from reconnaissance to launching a fatal strike—from an average of 4 weeks to as little as 2 days.

Figure: Bain estimate: AI-driven attack chain time compressed from an average of 4 weeks to as little as 2 days

Faced with adversaries’ asymmetric attacks, enterprise security leaders are eager to significantly increase the frequency of in-house offensive and defensive testing. As many as 83% of respondents believe that substantially increasing security posture testing frequency is critical, and a full 54% of executives view it as a non-negotiable lifeline for business survival.

As one CISO at a multinational high-end manufacturing giant said: “Facing AI-enabled new threats, we must find vulnerabilities faster than hackers do, before they knock on the door.”

Figure: Survey of 100 CISOs: More than 80% of respondents highly prioritize increasing security posture testing frequency

The advanced hands-on offensive and defensive experience accumulated by professional human penetration testers remains an indispensable core asset for any modern enterprise. However, it is undeniable that the scope of systems and applications that can be covered by purely manual human inspection is extremely limited.

For this reason, most enterprises are decisively breaking the old habit of conducting only one or two sample penetration tests per year, instead treating routine manual attack-and-defense exercises as a baseline floor, while layering on AI red team systems capable of 7x24-hour uninterrupted autonomous discovery.

The survey results clearly confirm this future trend: today, only 18% of surveyed companies can ensure that offensive testing runs continuously and automatically after every code release; but over the next three years, as many as 83% of companies are expected to incorporate continuous automated red-and-blue adversarial testing into standard enterprise infrastructure.

Figure: Projected increase in offensive security testing frequency: continuous automated testing will reach 83% within three years

This shift in technology paradigm is profoundly reshaping the lifecycle of software build and operations. Every code build and deployment will automatically trigger in-depth attack-and-defense simulations, allowing R&D engineers to directly see real vulnerability attack-and-defense path maps hours before production launch.

Figure: Continuous offensive testing deeply integrated into modern CI/CD and software release pipelines

Finding vulnerabilities is important, but automated remediation is the decisive factor.

In the vulnerability discovery process, simply finding vulnerabilities is often the easiest part. What truly burdens enterprise security and engineering teams is the prolonged struggle to drive large volumes of vulnerabilities to full remediation.

Remediating a deep vulnerability means security teams must identify the corresponding business owners across hundreds or thousands of microservices, determine exactly which code or configuration files need to be changed, and conduct repeated, time-consuming manual testing to ensure that applying a patch will not unexpectedly disrupt core production services.

As a security leader at a well-known fintech giant stated, with existing AI scanning tools, you can indeed find a large number of vulnerabilities very quickly, but your engineering team cannot remediate them at the same speed.

A Validated Attack Path helps engineering teams greatly prioritize remediation. It clearly and intuitively shows whether an attacker can truly reach the vulnerable node, which security controls effectively block the path, and which core configuration changes can reduce the largest risk exposure with the least development effort.

Figure: Validated attack paths greatly help engineering teams focus on closed-loop remediation of core risks

An executive at a leading global cybersecurity platform summarized this generational opportunity as 'Fix It With The Find It'. This means future security products cannot stop at issuing a lengthy alert list; they must directly generate tested remediation code Pull Requests with precise contextual dependency explanations.

Startups that seamlessly integrate vulnerability discovery, automated attack chain validation, and automated patch remediation into a closed loop will hold the most critical commercial positions in this industry wave.

SOC Operations Expansion: Automating One Specific Action at a Time

As external attack tempo is compressed to extreme levels, security operations center (SOC) analyst teams must investigate and respond with containment and blocking faster. However, the ceiling on enterprises' ability to maintain high-level defenses has long been constrained by a severe shortage of Tier-1 and Tier-2 analysts.

AI-native SOC tools are taking over tedious, repetitive basic tickets. Autonomous agents can correlate evidence chains scattered across endpoint, network, and identity logs in milliseconds and systematically complete standard routine checks, allowing highly skilled human analysts to focus on the most threatening advanced stealth attacks.

However, in practical deployment, how much system action authority to grant these AI agents is a red line that every prudent CISO must seriously weigh. Survey data shows that 36% of respondents are comfortable letting agents autonomously conduct complex investigations and evidence forensics, but only 13% of CISOs are willing to let agents fully autonomously execute blocking and code patch changes in production environments.

Figure: Survey of 100 CISOs: willingness to grant AI agents autonomous authority across SOC response actions

Searching correlated data in underlying read-only logs is fundamentally different in risk from directly applying firewall policies or disconnecting server networks in production systems. Therefore, most security teams currently prefer to fully delegate complex, lengthy forensic investigation workflows to agents while firmly retaining human confirmation (Human-in-the-Loop) for high-risk blocking actions.

This is also the core reason a16z strategically invested in Cotool. Cotool enables enterprise security teams to quickly package their proprietary investigation expertise and SOP guidance into autonomous, collaborative agent swarms. These agents can seamlessly span existing toolchains to triage alerts, conduct enterprise-wide investigations, and intelligently draft response plans, tackling one core workflow at a time while maintaining security risk controls and unlocking substantial productivity for SOC teams.

Safeguarding AI assets and agent security within the enterprise

AI technology has long been deeply embedded in the core daily business workflows of modern enterprises. Survey data show that as many as 92% of surveyed large enterprises have employees using generative AI tools in production-grade daily work; 81% are conducting in-depth internal testing or formally operating autonomous AI agents; and 80% have deployed customer-facing generative AI interactive products.

Chart: Survey of 100 CISOs: Production-grade penetration statistics of generative AI and autonomous agents within enterprises

These highly empowered internal systems typically have unprecedented privileges: they can directly access core code repositories, retrieve all non-public confidential corporate documents, and even act on behalf of users across critical business SaaS platforms. As one security executive at a large financial services institution observed: 'Today, every employee's desktop is evolving into a high-density micro data center.'

The resulting internal security concerns are highly pertinent. In the survey, 47% of CISOs identified employees leaking highly confidential internal data to uncontrolled third-party AI tools as their greatest fear; 41% of respondents worry that highly privileged internal AI agents could perform unauthorized privilege escalations or suffer malicious prompt injection (Prompt Injection) attacks, leading to systemic compliance failures.

Chart: Survey of 100 CISOs: Top security risks of concern when deploying internal AI assets

For a significant portion of large enterprises, the unresolved security risks above have objectively slowed business innovation. As many as 63% of surveyed CISOs acknowledged that security and compliance considerations have delayed, to varying degrees, the full launch of at least some internal AI pilots.

Security teams urgently need a comprehensive radar hub: real-time visibility into which AI tools and agents are running on the corporate network, who owns them from a business perspective, what system access credentials they have been granted, and what actions they are actually executing moment by moment.

However, precisely observing these micro-level actions poses far greater technical challenges than imagined. Modern autonomous agents often run directly on employees' local computer endpoints and fully inherit users' elevated access permissions. For traditional security monitoring systems built around network boundaries or cloud APIs, the operations initiated by agents look no different from legitimate mouse clicks by the employee.

This is the underlying logic behind our significant investment in Neo. Neo's founding team previously led development of the core engine at endpoint security leader SentinelOne. They are now rebuilding the endpoint security defense layer from the ground up for a new paradigm driven by large fleets of AI agents, precisely capturing every agent action and creating a robust barrier between compliance security and innovation efficiency.

Risks targeting 'people' are undergoing an AI-level evolution.

Although underlying vulnerabilities and machine-driven attacks are becoming increasingly sophisticated, the weakest and hardest-to-defend link in cybersecurity remains human beings themselves. Ironically, '123456' still consistently ranks first among the most common weak employee passwords at global enterprises.

As AI matures, it is pushing traditional social engineering phishing and identity fraud targeting human psychology into a terrifying dimension of industrialized, high-fidelity forgery. A forged phishing email can mimic a colleague's tone with no discernible difference; a voice call can clone a direct supervisor's voice and breathing in real time; even a participant in a remote video conference can convincingly appear to be a familiar team executive.

When Adaptive Security founder Brian Long visited the a16z office for a fundraising pitch, he showed all the partners present two seemingly identical recent photos of Elon Musk and asked everyone to guess which one was the unaltered, authentic photograph.

Chart: Adaptive Security phishing and anti-forgery field test: authentic photos and AI deepfakes (Deepfake) are difficult to distinguish

Shockingly, the vast majority of the seasoned technology investors present got it wrong.

This scene vividly reveals the identity-trust crisis facing enterprises today: modern social-engineering impersonation has long moved far beyond the primitive stage of grammatical errors and odd sender domains. Even a core employee trained in strict confidentiality and highly vigilant day-to-day can see their psychological defenses shattered in an instant by millisecond-seamless spoofed voice and face attacks.

This moment has strengthened our conviction that traditional static employee compliance quizzes and crude phishing simulations must undergo a full AI revolution. Adaptive uses generative large models to create highly realistic, context-aware dynamic deepfake exercises for enterprise employees in real time, fundamentally rebuilding employee resilience in the AI era.

A more severe challenge lies not only in the realism of the impersonation, but in the explosion of attack scale once the marginal cost of forgery approaches zero. Attackers need only type a single command to generate, concurrently and instantly, thousands of tailored spear-phishing variants customized for employees across different departments, overwhelming traditional defenses based on keywords and blocklists.

Massive volumes of code are being generated, and the bar for code security is rising across the board.

In complex software engineering supply chains, even if you have full confidence in the code written by your own team, a single fatal backdoor in an open-source component or third-party service you depend on can cause the entire security edifice to collapse in an instant. The major supply-chain hijacking incidents recently disclosed are a stark wake-up call to this fragile reality.

Code security is not a new market. For decades, mid-sized and large enterprises have invested heavily each year in static application security testing (SAST) and software composition analysis (SCA) tools. However, with the widespread adoption of AI-assisted coding tools such as Cursor and GitHub Copilot, the productivity leverage of software development has been fully unlocked, and the volume of new code in enterprise codebases is multiplying at a geometric rate.

Huge volumes of code are arriving, while internal security engineering headcount cannot possibly double in step. As a result, nearly half of the 100 CISOs we surveyed (48%) firmly identify AI code security as one of their Top 5 core priorities for 2027, a metric that has nearly doubled compared with previous years.

As the CISO of a large financial services institution candidly put it: "The number of R&D engineers producing business code at their keyboards every day far exceeds the small number of security staff responsible for secure code review."

Demand is rising, but platform incumbents still enjoy first-choice priority.

Although defense priorities are undergoing a generational leap and new enterprise procurement budgets are being released in batches, in the actual procurement decision pipeline, established platform giants still enjoy dominant first-round inquiry priority.

When asked who they would approach first to deploy new AI security capabilities in their enterprise, as many as 44% of CISOs habitually cast their first vote for an existing platform vendor they have already purchased, hoping to meet the need through seamless feature upgrades to their existing suite.

As a Fortune 500 executive put it bluntly: "I trust CrowdStrike. It has an unmatched top-tier security brand reputation and battle-tested credibility."

However, the first call does not necessarily mean the final order. When we asked further what key trigger would lead security decision-makers to bypass established giants and buy from emerging startups, the survey data delivered a highly revealing business answer.

Chart: Survey of 100 CISOs: Core drivers pushing enterprises to purchase startup products beyond incumbent giants.

A full 61% of CISOs explicitly stated that if a startup vendor offers a new critical capability entirely absent from incumbent giants' product lines, they will decisively open a dedicated budget to purchase it; at the same time, 49% of buyers cite the overwhelming superior performance (Superior Performance) demonstrated by startup products as the decisive breakthrough point.

In most real-world scenarios, large enterprises buy cutting-edge security products from startups not to completely overturn or replace their existing platform foundations, but to precisely graft a sharp specialized blade alongside their current defenses and fill the technical gaps in high-risk trenches that traditional suites cannot cover.

The proliferation of market noise is making selection decisions even more difficult.

Looking at the current enterprise AI security ecosystem, hundreds of emerging startups have sprung up, broadly claiming to offer fully autonomous intelligent defense, millisecond-level alert tracing, and one-click closed-loop remediation that frees up human resources.

The exhibitor landscape at the premier Black Hat 2026 security conference vividly illustrates this intense competition: some narrow AI security subsegments contain as many as 43 competing vendors, resulting in severe homogenization of marketing concepts and positioning confusion.

Amid heavy market noise and sales bombardment, enterprise buyers are forced to spend significant time verifying claims: what percentage of grunt work can each product truly complete unattended? How much cleanup will human analyst teams still need to perform? What operational permissions do agents have in production environments? And if a false positive occurs, can the system provide a clear decision tree that supports forensic-level auditing?

The CISO of a leading global apparel brand vividly described his team's exhausting screening process: after carefully selecting from seven or eight similar emerging competitors, they ultimately chose only three for in-depth, hands-on on-site testing. In his words: 'All products look flawless in slide PPT demos, but only when you connect them to messy real business systems can you see who is a mule and who is a horse.'

The challenge is compounded by the fact that the technical baselines used for horizontal comparison are themselves evolving rapidly. New founding teams keep entering the space, existing products' capability boundaries change month after month, and every leap in the inference performance of underlying large foundation models puts the technical moats of existing products through a harsh retest.

As another enterprise security leader expressed a common sentiment of caution: 'I absolutely do not want to endlessly procure a pile of isolated, uncoordinated temporary single-point tools (Point Products) inside my company.'

However, it is important to recognize that this objective caution and restraint has not dampened strong demand for truly hardcore security technology; rather, it coexists deeply with robust purchasing momentum. Security teams at major multinational enterprises are acting with unprecedented urgency, actively applying for dedicated procurement budgets, advancing pilot programs, and formally integrating the first wave of field-tested pioneering products into everyday production workflows.

Forward-Looking Outlook

In our memory, no technology wave has ever reshaped the overall cybersecurity agenda of top enterprises in the United States and around the world as thoroughly and profoundly, and at such overwhelming speed, as AI has.

Budgets are being raised significantly, teams are expanding headcount, and the vision and scope of defense teams are no longer confined to the security comfort zone defined by traditional platform suites.

This is a golden historical window for great cybersecurity startups to emerge. In the new frontiers of cutting-edge AI offense and defense, agent security auditing, and autonomous remediation, bloated legacy incumbents have not fully caught up, while forward-looking enterprise security leaders are already opening their doors, eager to bring stronger new forces into the deepest trenches.

We firmly believe that, alongside this unprecedented revolution in offense and defense paradigms, the next generation of cybersecurity giants is being rapidly incubated in this fertile ground. They will build the grand defensive line that safeguards global digital civilization for decades to come, centered on the new security capabilities that are only beginning to emerge in enterprises.

Special thanks: We thank Zane Lackey, Joel de la Garza, and David George for their outstanding contributions to the in-depth research and analysis in this article.

Disclaimer: All content in this long-form article is for informational reference only and must not be treated as a formal basis for any legal, business, investment, or tax decision. Some investment targets mentioned in the article are portfolio companies backed by a16z funds; for the complete portfolio and core interest disclosures, please refer to the official site.

Source: TechFlow