As copilots give way to autonomous agents, “security for AI” is emerging as cybersecurity’s next big category, and the industry’s channel ecosystem is being asked to build it
There are two conversations about AI in cybersecurity, and they are often treated as though they are the same.
The first is AI for security: using AI to improve detection, investigation, vulnerability analysis, analyst productivity and response. The second is security for AI: protecting the models, data, applications, identities, APIs and increasingly autonomous agents that enterprises are putting into production.
The second conversation is younger. It may also become one of the most important cybersecurity markets of the next decade.
2022 – The warning
In 2022, the AI discussion in mainstream cybersecurity research was still primarily about attackers gaining AI capability and defenders using AI to improve human productivity. McKinsey warned that attackers would use AI, machine learning and automation to compress the attack life cycle from weeks to days or hours, while organisations should use automation for lower-risk, repetitive work and free security teams for higher-value decisions.
What is striking in retrospect is what was not yet at the centre of the discussion: the security of enterprise AI systems themselves. That gap shows how quickly the market moved.
2024 – The evidence
By 2024, McKinsey was treating “securing AI” as a stand-alone cyber market segment. Its survey found enterprises prioritising capabilities such as model observability, AI governance, sensitive-data scanning, production vulnerability monitoring, code scanning, protection against input manipulation, data-poisoning mitigation, continuous red teaming and model-drift monitoring.
More than 97% of surveyed customers expected to increase spending with outside vendors to secure AI use cases. Just as important, customers said they would initially give existing vendors the opportunity to solve these problems but would look elsewhere if incumbents could not meet the requirement. In McKinsey’s survey, 58% said they would seek solutions already available in the market rather than simply wait for an existing supplier to build the missing capability.
That is a powerful signal for the channel. A new AI-security company does not have to replace an established cybersecurity platform to become strategically important. It only has to solve a meaningful risk that the existing architecture does not yet address well enough.
2026 – The reality
The World Economic Forum’s Global Cybersecurity Outlook 2026 shows how decisively the issue has moved from awareness to action. Ninety-four percent of survey respondents expect AI to be the most significant force changing cybersecurity, and 87% identify AI-related vulnerabilities as the fastest-growing cyber risk during 2025. The proportion of organisations with processes to assess AI-tool security almost doubled from 37% to 64% in one year.
The nature of the concern is also changing. WEF reports that data leakage associated with generative AI is now a leading concern, at 34%, ahead of concern about advancing adversarial AI capabilities at 29%. In other words, security leaders are no longer worried only about what attackers can do with AI. They are increasingly worried about what their own use of AI may expose.
Verizon adds another signal: frequent employee use of unapproved “shadow AI” tools rose from 15% to 45% in a single year in its dataset. The enterprise is adopting AI faster than traditional approval, governance and security processes can always keep up.
AI security is not one problem
It is tempting to imagine AI security as another product category that can be added to a technology stack. In practice, it spans several existing domains and introduces genuinely new ones.
Data- AI systems consume information. Security teams need to understand what data can enter prompts, what confidential information may be retained, where model inputs and outputs are stored, whether sensitive information can leak, and whether the data used to ground or train the system is trustworthy. Compliance sits alongside security: AI use must meet data-protection, data-residency and sector-specific regulatory obligations, and organisations need to be able to demonstrate how data has been handled.
Model- Models can be manipulated, extracted, poisoned or induced to behave in unintended ways. Model drift and quality changes can create operational risk even without a conventional attacker. AI governance and model-risk controls therefore begin to intersect directly with cybersecurity.
Application- Most enterprises do not deploy a raw model; they build applications around it. Those applications still have code, software dependencies, APIs, authentication mechanisms and traditional vulnerabilities. AI security extends application security rather than replacing it.
Identity and privilege- Who is allowed to use the AI? What enterprise data can that session access? Which external systems may the application call? What happens when an AI tool operates with a highly privileged service identity?
Agents and agency- This may be the most consequential layer. A chatbot produces an answer. An agent may query a database, invoke an API, send a communication, modify a record or initiate a workflow. Once software can act, the security question shifts from controlling information to controlling authority.
|
The defining question of agentic security may become: how much authority should a reasoning system possess, under what conditions, and how quickly can that authority be withdrawn? |
Shadow AI is an operating-model problem
The instinctive security response to unapproved AI is to block it. Yet employees turn to these tools for the outcomes they produce – faster proposals, better analysis, software productivity, customer-service assistance, workflow automation or easier access to knowledge.
The history of Shadow IT is instructive. Technology decisions gradually moved into individual business functions because those functions understood the outcome they needed. Shadow IT has long since transformed into “Sanctioned IT” across functions and departments, while central IT shifted toward architecture, governance, integration and enablement. AI is likely to accelerate the same pattern. Every function can now discover a tool that promises a local productivity gain.
The security answer therefore cannot be prohibition alone. A mature AI-security model needs discovery of actual AI usage, data classification, policy, identity and access control, approved alternatives, model and application assessment, monitoring, education and response. The organisation needs to make the safe path easier than the unsafe path.
The AI-security market will be an ecosystem market
No single technology category naturally owns every layer of AI security. Identity vendors will participate. Data-security vendors will participate. Cloud providers, application-security companies, governance platforms, SecOps vendors and specialist AI-native security firms will all contribute. Established cybersecurity companies will extend their platforms, while new firms will appear around problems that did not exist at scale a few years ago.
This is exactly the kind of market in which discovery and orchestration matter. Partners and customers cannot continuously evaluate every new AI-security company globally. New vendors may have strong intellectual property but limited market access. The value of an ecosystem is its ability to evaluate, contextualise, integrate and take credible innovation to market.
Next – One step ahead
Companies that see AI security as both a technology opportunity and an architecture challenge, with an objective not to create a separate island called “AI security” and ignore everything around it. It is to connect AI security with cloud, identity, data, application security, governance, operations and services – then help partners build the capability to take that architecture to customers.
For the last few years, the industry asked how AI would change cybersecurity. The more urgent question now is how cybersecurity must change because AI has become part of the enterprise itself.
By- Sridhar S, President – SSG, India, Redington Limited
Source notes for editorial validation
For fact-checking; remove if the outward-facing format does not carry references. 1. McKinsey & Company, “Cybersecurity trends: Looking over the horizon”, 10 March 2022. (Also summarised in McKinsey, “What is cybersecurity?”, April 2023, user-supplied source.) | 2. McKinsey & Company, “The cybersecurity provider’s next opportunity: Making AI safer”, November 2024. (User-supplied source.) | 3. World Economic Forum, “Global Cybersecurity Outlook 2026”, January 2026. | 4. Verizon, “2026 Data Breach Investigations Report (DBIR)”, 2026.
