Keep pulling the thread on Bradon Rogers.
AI agents can be exploited by following hidden instructions buried in an application, which they mistake for commands from a human user.
A significant portion of AI development is consumer-driven, which creates security and compliance concerns for large enterprises like financial services firms and healthcare providers when these tools are used in their environments.
Island recently announced a series of new capabilities focused on enabling the safe use of AI within the enterprise.
The rapid adoption of consumer AI tools in the enterprise makes it more difficult for organizations to adhere to their regulatory requirements and data protection needs.
Island's platform enforces policies to prevent corporate data from being used in personal AI tenants, such as a personal Gemini account, while still allowing employees to use those personal tools.
Analyst firm Gartner has published research advising enterprises to be cautious and temporarily avoid using agentic browsers due to security risks.
Island's AI Protect product uses a service called Island Desktop to assert policy and govern the usage of local AI applications outside of the browser.
Traditional security providers that inspect traffic at upstream cloud pinch points often must resort to blocking personal AI services entirely, as they cannot effectively distinguish between personal and corporate tenants.
Island's technology can govern modern AI-centric protocols, including MCP (Machine-to-Machine Communication Protocol) calls, to ensure agentic workflows operate within a permissible scope.
There are over 18,000 browser extensions classified as AI, creating a significant potential security risk for enterprises.
Island's AI Publish feature allows enterprises to securely deliver and apply policies to applications built with "Vibe coding" tools like Lovable, Cursor, and Claude Code.
By observing user workflows, Island's platform can proactively suggest relevant prompts or automated workflows to users, such as reducing a five-minute task across five systems to 30 seconds.