Digital Content Editor, Eve Goode spoke with Michael Leland, VP and Field CTO at Island who discussed how generative AI is exposing a blind spot in traditional SASE architectures and why protecting what users do is becoming as important as securing where they connect.
How is generative AI exposing gaps in traditional SASE architectures?
AI is highlighting the unsuitability of network-based security architectures in a world of work that increasingly relies on SaaS applications and desktop AI tools like Claude, Cursor, Codex and numerous other cowork applications.
Traditional Secure Access Service Edge (SASE) examines traffic routed to a cloud point of presence (PoP), where the traffic is broken down, inspected and forwarded.
That model made sense when most work happened on managed devices connected to corporate networks.
Security teams could understand a lot about risk by simply analysing network traffic alone. This is yesterday’s approach.
Today, employees rely on a variety of applications and AI tools, including AI agents that access data and execute workflows at a scale no human can match.
Modern encryption can also make traffic from many AI tools difficult or even impossible to inspect.
That fundamental incompatibility forces legacy SASE administrators to create bypass lists.
Yet every exemption they make risks everyday work sessions landing outside the security perimeter.
End users also feel the impacts. When SASE enforcement requires everything to be routed through a distant proxy, performance suffers with lags and broken SaaS workflows.
Users themselves then seek workarounds to stay productive, creating further blind spots that expand the attack surface.
AI has exposed this gap in a way that’s difficult to ignore.
If security creates friction, then people will work around it. That’s not a user problem, but an architectural one and organisations are responding in the only way SASE permits – to block AI or allow it without the required controls, checks and balances.
Why is visibility into what users do becoming as important as where they connect?
For years, security was centred around understanding who a user was, the device in use and the connected applications.
These remain important foundations of any security strategy, but organisations now need additional insight into what employees are doing – both on the web and at the desktop level.
A staff member might be using an approved device to access a permitted AI tool, but security teams also need to check that they are not uploading confidential or sensitive information or moving data between corporate and personal environments.
Simply knowing that an individual is connected to ChatGPT or Microsoft Copilot isn’t adequate.
Here, legacy SASE again presents a barrier because data uploads, prompts and AI-generated outputs all occur at the point of intent, not in transit.
Traditional SASE-based network inspection can see
connections, but not clipboard actions, prompt content, local file system access, or file transfers between all apps. And that is a problem. If you can’t see what’s happening within AI sessions, you can’t effectively govern it.
Why is the browser emerging as a critical security control point for AI?
Most SaaS applications and AI tools run or are accessed through the browser, meaning it is often the place where files are uploaded, prompts are written, content is copied and data is shared.
Browser-based activity happens before information becomes network traffic.
A network proxy can see that a connection was established with an AI service, but the user interaction that occurred within it can then only be seen within the browser itself.
Therefore, improving browser-based controls will become particularly important as AI and SaaS adoption accelerate.
Organisations need to be able to govern AI without preventing employees from using it.
To do that, security teams need visibility into the prompts, uploads, downloads and data movement within AI and SaaS tools.
This is insight that the browser is uniquely positioned to provide, being the place where work now often happens.
But governance must extend beyond the browser, too.
End users are doing more work on AI desktop apps, requiring common policies to manage and audit all AI actions wherever they occur.
How can organisations strengthen AI governance without replacing existing SASE investments?
SASE remains an important part of security architecture, supporting organisations with aspects of network security such as zero-trust access and identity-based controls.
Yet organisations now also need to see how users are using SaaS applications, the agents they might invoke and AI tools accessed both in a browser and from the desktop.
There are several controls that can be used to ensure AI access and usage align with company security policy, without forcing a choice between productivity and protection.
For example, data boundaries define which AI tools and workflows are approved for organisational use, ensuring data doesn’t cross into unsanctioned territory.
Content-aware detection solutions can also inspect what users send to AI in real-time, from pasted text to uploads, potentially blocking a user approved to use AI for general research from pasting sensitive data into the tool.
Meanwhile, AI visibility and control extensions can give IT teams real-time insights into risk, showing what employees access, what they send and where outputs go.
And last-mile policy enforcement can then restrict or redact sensitive content at the point of submission, while keeping AI outputs within approved workflows.
Runtime AI policies can enforce data protection, content moderation and security checks on all prompts, responses, tool calls and tool call results – providing a high-fidelity audit trail and governance across all modalities of AI.
What will the next generation of SASE need to deliver in an AI-driven workplace?
New SASE architecture must appreciate that AI agents, SaaS applications and remote work have become the new norm for how work gets done, moving enforcement to where work happens.
Traffic must be analysed, inspected and secured at the most appropriate location: the user experience layer inside the browser, via applicative hooks into desktop AI apps, within a Model Context Protocol (MCP) gateway to enforce least-privilege access and NHIs, or at the LLM gateway
Legacy SASE penalises the end user, resulting in slower applications.
On unmanaged devices, it often doesn’t work at all, requiring agents and certificates that contractors and third parties can’t install.
New SASE architectures will need to address these shortcomings, making work faster and smoother for everyone, on any device.
Direct-path traffic will be important, meaning applications load faster.
Unnecessary Transport Layer Security (TLS) break and inspect must also be addressed to mitigate broken session and certificate errors.
Deployment must be achievable in minutes rather than months, across both managed and unmanaged devices.
The future of SASE will be defined by its ability to support AI and SaaS usage while maintaining control, giving employees and AI agents the freedom to work faster.
All while ensuring that governance, security and accountability remain built into the workflow itself.
Michael Leland is a VP and Field CTO at Island with 30+ years in networking, operations and cybersecurity. In previous roles at SentinelOne, McAfee and NitroSecurity, he shaped XDR and identity security strategy, drove technical go-to-market direction and built out a SIEM technology vision that led to NitroSecurity’s acquisition by McAfee.