Check Point Report Finds 51-Point Gap Between AI Security Strategy And The Architecture To Enforce It

Enterprises have updated their security strategies for the AI era but lack the infrastructure to back them up, according to new research from Check Point Software Technologies that reveals a widening disconnect between intent and capability.

The Check Point 2026 Cloud Security Report, based on a survey of 1,042 cybersecurity and IT professionals conducted by Cybersecurity Insiders, found that 77 per cent of organisations have changed their security strategy in response to AI adoption. But only 26 per cent report having the architecture to enforce those changes.

That 51-point gap between strategy and execution is already producing consequences. More than half of the organisations surveyed – 54 per cent – reported confirmed AI-related security incidents over the past year, while another 24 per cent could not determine whether they had been affected due to limited visibility. Taken together, 78 per cent have either experienced a confirmed incident or cannot rule one out.

AI has moved into production faster than security has followed

The report paints a picture of AI that has moved well beyond experimentation. Seventy per cent of organisations now run generative AI workloads in production, with 10 per cent describing AI as core to their business operations. At the same time, 64 per cent have deployed AI agents in pilot or production environments, including 12 per cent that have granted those agents privileged access to core systems.

Yet 83 per cent of respondents indicated that securing generative AI is harder than securing traditional applications. The security challenge has shifted from governing what employees ask AI tools to do toward governing what AI systems are permitted to do on their own.

Paul Barbosa, Vice President of Cloud Security and SASE at Check Point Software Technologies, pointed to the structural nature of the problem.

“AI adoption has outpaced the architecture built to govern it,” Barbosa noted. “Agents are acting inside live systems; data is moving through external AI services, and most enterprises still lack the visibility and enforcement to keep pace.”

“Visibility, control and security need to be present at all layers in the stack AI workloads will operate in,” he added.

Visibility remains a fundamental weakness

The report identified visibility as a foundational gap. Only 5 per cent of organisations reported full visibility into which AI tools employees are using, what data those tools access and where that data goes once it enters an AI workflow. That leaves 95 per cent making policy and enforcement decisions based on an incomplete picture.

The problem extends beyond discovery. A separate finding showed that only 5 per cent of security tools can reliably distinguish legitimate AI activity from suspicious or unauthorised usage. Browser-based AI assistants may leave little endpoint evidence, API calls from scripts may not appear in SaaS discovery and AI agents using existing service accounts can look like normal system activity.

Traditional discovery tools were designed to find known applications, managed endpoints and catalogued SaaS usage. AI activity frequently bypasses those assumptions.

Shadow AI and employee workarounds are widespread

The gap between strategy and enforcement is showing up in user behaviour. Forty-two per cent of organisations reported that employees bypass AI security controls when those controls slow them down. Workers paste sensitive data into personal AI accounts, use browser-based tools their endpoint agents do not cover or take whatever path feels faster than the approved one.

Only 14 per cent of organisations have formal generative AI policies that are actively enforced and audited. Meanwhile, 51 per cent of security teams believe they are enabling AI adoption, yet 21 per cent acknowledge that slowing AI adoption for security reasons has already cost their organisation competitive ground.

The report found no dominant approach to governing employee access to AI tools. Twenty-four per cent have no AI-specific access controls at all, 22 per cent rely on endpoint security only, 19 per cent apply different rules depending on whether users are on or off the network, 19 per cent block external AI tools outright and only 16 per cent enforce the same controls consistently regardless of location.

Infrastructure was not built for AI traffic patterns

AI is changing network traffic in ways that existing security architectures were not designed to handle. Fifty-one per cent of organisations reported more API-driven traffic, 48 per cent saw increased traffic to external AI services, 42 per cent noted new user-to-AI communication flows and 33 per cent reported less predictable traffic patterns overall.

Only 24 per cent of respondents indicated their current network security tools can fully inspect AI traffic without degrading application performance. The remaining 76 per cent face inspection gaps, performance trade-offs or limited confidence in their controls.

At the same time, 67 per cent of organisations report fragmented security policies across their hybrid environments, and 64 per cent indicated their architecture needs moderate or significant redesign to support AI workloads.

Datacentre perimeter security falls short

The infrastructure challenge extends to on-premises environments. Twenty-nine per cent of organisations are already migrating AI training and inference workloads to their own datacentres, with another 49 per cent planning or considering the same move.

Seventy-six per cent rate datacentre perimeter security as critical for protecting AI workloads, but only 35 per cent say their current perimeter is capable of handling them – a 41-point gap between perceived importance and actual capability.

The broader deployment picture reflects this complexity: 52 per cent of respondents indicated their AI models and AI-powered applications run across a hybrid mix of cloud, on-premises and third-party services.

Web application firewalls are struggling with AI traffic

At the application layer, existing controls are showing strain. Only 22 per cent of organisations rate their current WAF or WAAP tools as effective against generative AI-specific attacks such as prompt injection. Thirty-six per cent described their tools as only somewhat effective, while 28 per cent indicated their current tools were not designed for AI threats.

The operational impact is visible: 71 per cent reported increased WAF false positives since generative AI adoption began. Traditional WAF logic was tuned for human-driven web traffic and known attack signatures. Generative AI applications introduce long prompts, streaming responses, model-specific APIs and service-to-service interactions that do not map to those assumptions.

Coverage of AI assets is also uneven. Only 38 per cent of organisations reported that their application security controls protect public-facing AI web applications, while 32 per cent protect APIs connecting to AI services, 28 per cent protect LLM endpoints and 20 per cent protect AI agents. Twenty-one per cent reported no AI-specific application security at all.

Runtime controls lag behind deployment

Inside AI applications, the exposure is more direct. Models read sensitive data, generate outputs and invoke tools and APIs in real time. Only 17 per cent of organisations have broadly deployed runtime LLM controls such as input validation, output filtering and tool-use authorisation.

Fifty-six per cent have no formal security testing process for generative AI applications, or test only on an ad-hoc basis. Many AI applications are reaching production without structured prompt-injection testing, without repeatable adversarial evaluation in the CI/CD pipeline and without consistent retesting when models or workflows change.

Non-human identity emerges as a top concern

AI is creating access paths that enterprises were not built to govern. Forty-eight per cent of organisations ranked managing non-human identities – AI agents, APIs and service accounts – as their leading AI-related identity challenge. Forty per cent cited API key and credential sprawl, 37 per cent pointed to over-privileged service accounts and 34 per cent flagged difficulty auditing AI actions.

The AI supply chain presents further exposure. Forty-six per cent of organisations have no structured security assessment process for AI vendors, and only 7 per cent scan AI models for vulnerabilities or malicious code before deployment.

Data flows lack governance

Data movement represents one of the broadest areas of AI security exposure. Forty-four per cent of organisations cannot trace where sensitive data goes once it enters an AI workflow, and only 7 per cent have full visibility into data lineage through AI processing.

Twenty-five per cent of organisations permit source code in generative AI tools, potentially exposing proprietary business logic and internal configurations to external systems with limited exfiltration control. Only 15 per cent have deployed data loss prevention controls specifically configured for AI data flows.

Detection exists but real-time prevention does not

Across prompts, data flows and outputs, the report found the same pattern: organisations can often detect AI-related threats but rarely block them in real time.

Only 13 per cent can block a malicious prompt before it reaches a model. Only 16 per cent can block sensitive data from reaching AI services in real time. And only 5 per cent can reliably block unsafe AI-generated content before it reaches users or downstream systems.

The report described this as “detection without prevention” – a model that produces evidence after the event rather than control at the moment it matters.

Complexity is increasing, governance is fragmented

Eighty-eight per cent of organisations reported that AI adoption has increased the operational complexity of their security programs. The top barriers cited were lack of visibility at 28 per cent, skill gaps at 25 per cent and disconnected tools at 16 per cent.

Accountability for AI security remains spread across multiple roles. Forty-four per cent assign it to the CISO, 40 per cent to cross-functional committees and 36 per cent to CIO or IT leadership. Only 14 per cent have a dedicated AI security leader.

The governance execution gap is visible in the numbers: 45 per cent have documented AI security policies, but only 14 per cent actively enforce and audit them. On regulatory preparedness, only 7 per cent described themselves as fully prepared for AI-specific requirements such as the EU AI Act and NIST AI Risk Management Framework.

Organisations are consolidating their security approach

The market is responding by moving toward broader, more unified operating models. Thirty-seven per cent of organisations are actively consolidating around existing platform vendors, nearly double the 20 per cent investing in best-of-breed point solutions. Eighteen per cent are taking a wait-and-see approach, 13 per cent are building internally and 12 per cent are pursuing a service-led model.

Eighty-six per cent rated unified security management across datacentre, cloud and edge as critical for AI workloads. Shadow AI discovery leads investment priorities at 31 per cent, followed by AI-powered SOC automation at 25 per cent and adversarial defence capabilities at 19 per cent.

The 2026 Cloud Security Report was produced by Cybersecurity Insiders with the support of Check Point. The survey gathered responses from 1,042 cybersecurity and IT professionals across industries including technology, financial services, healthcare, manufacturing and government. A stratified sampling approach yielded a 95 per cent confidence level with a margin of error of plus or minus 3 percentage points.

Check Point Software Technologies provides cyber security solutions to more than 100,000 organisations worldwide.

Last Updated on May 27, 2026 by Nick Ross