Cloudflare’s head of SASE, Dave Barnett, described an AI strategy built around three goals: running inference near users, securing workplace use of AI and understanding how attackers may apply the same technology. The approach connects performance with security rather than treating them as separate infrastructure decisions.
Cloudflare’s global edge network gives it a natural role in low-latency inference, while its zero-trust products can govern access to models and data. The central proposition is that AI should be close enough to respond quickly but controlled enough to prevent convenient tools from becoming an uncontrolled path for sensitive information.
Edge Inference Reduces the Distance to Users
Cloudflare introduced Workers AI to let organisations and individuals run inference on its distributed network using NVIDIA GPUs. Training would generally happen elsewhere, but deployed models could execute near the people or devices making a request. Image recognition was one example where shorter network distance can support a more responsive experience.
Latency matters for interactive AI. A small delay may be acceptable for a long report, while voice, vision and security applications require rapid responses. Processing near users can also reduce the need to route every request to one central region, improving resilience and potentially supporting local data requirements.
Edge deployment introduces operational challenges. Models and software must be updated consistently across many locations, while monitoring needs to provide a coherent view of performance. Developers should know which region handles a request, what data is retained and how the platform behaves when a nearby node lacks capacity.
Zero Trust Can Govern Workplace AI
Barnett connected AI adoption with data-protection requirements. Employees may paste confidential information into public tools or use unsanctioned services because they are convenient. A zero-trust approach verifies identity and context for each interaction, applying policy without assuming that a user or device is safe merely because it is inside a corporate network.
Cloudflare’s own zero-trust journey began with engineers who needed to reboot servers around the world. Replacing VPN-based access with a more usable system improved the employee experience and later produced Cloudflare Access. The lesson is important: security controls gain adoption when they reduce friction instead of forcing staff to work around them.
For AI, policy might restrict which models a department can use, prevent sensitive data from reaching an external service or log requests for investigation. Controls should remain proportional. Excessive surveillance can undermine trust, so organisations need clear purposes, limited retention and transparency about how employee interactions are monitored.
AI Strengthens Both Defenders and Attackers
Barnett’s third focus was whether malicious actors could use AI against defenders. Generative systems can help write convincing messages, adapt content and accelerate routine research. Security teams can use similar capabilities to summarise alerts, detect patterns and support analysts, creating an ongoing contest over speed and scale.
The answer is not to classify every AI-generated request as hostile. Detection should examine behaviour, identity, volume and context. Attackers can use ordinary automation without AI, while legitimate applications may generate unusual traffic. Reliable controls focus on observable risk rather than assumptions about which tool produced an action.
Human analysts remain responsible for consequential decisions. An AI system can prioritise an alert or suggest an explanation, but false positives can interrupt users and false negatives can expose data. Teams need evaluation based on real incidents and a feedback loop that improves performance without silently expanding automated authority.
Performance and Security Share an Architecture
Cloud-native infrastructure can apply traffic controls, identity policy and inference close to the same user. That integration may simplify deployment and reduce the gaps created when separate vendors handle networking, security and AI serving. It also concentrates dependency, making transparency and portability important for customers.
Organisations should test how the complete service fails. If an edge node, identity provider or model becomes unavailable, applications need predictable fallback behaviour. Security policy should not disappear during degradation, and an unavailable AI feature should not block essential work without an alternative route.
Cost is another design factor. Distributed inference can improve response time, but model size and request volume still shape economics. Developers should match the model to the task and measure completed outcomes rather than assuming that the largest model at the nearest edge is automatically best.
Data location should be visible to application owners. Running close to a user may support regional processing, but proximity alone does not prove compliance. Customers need configuration and documentation showing where prompts, logs and model outputs travel. Security teams should review the complete data path, including observability services and support access, before treating an edge deployment as local by default.
Cloudflare’s position illustrates a broader convergence. AI inference is becoming a network workload, and AI access is becoming an identity and data-governance problem. Organisations need both rapid delivery and controls that follow users wherever they work.
Barnett’s framework is useful because it includes benefit and abuse in the same picture. Edge compute can make AI responsive, zero trust can make authorised use safer and security research can prepare for malicious applications. The strongest implementation will make those layers reinforce one another, delivering speed without asking businesses to surrender oversight of their data or their users.
