Greywall AI Agent Sandboxing & Aligning Capability with Security
Greywall is a software-defined agent sandbox & proxy that was born out of a frustration with constantly having to choose between convenience and safety.
A presentation by Max, founder of Greyhaven, a Montréal-based startup dedicated to building sovereign AI systems.
In late 2024, I discovered my first AI coding agent (Aider), and I was immediately faced with an increasingly common decision: how much effort do I want to put into using this thing safely?
At the time, the FOMO fuel wasn't as potent as it is today, and I took the time to set up a docker container to test the thing out. I couldn't think of a plausible way that one of these language models could do much harm in a basic sandbox.
But it was immediately clear how powerful this paradigm was, and I found myself spending more and more effort to build automations around containers. My team built a CLI tool which automated docker commands, agent builds, MCP tools.
But our container CLI tool was clunky and difficult to introspect. Every day I felt like I was fighting a tradeoff between security and capabilty.
So we went back to the drawing board and thought, "what do we really need to feel comfortable using local AI agents?" And so we designed something fundamentally different: Greywall. A software-defined sandbox that acts as an intermediary layer between an application (e.g. an Agent) and the operating system resources.
This pattern allows for users of agent CLI tools to have fine-grained, runtime visibilty & controls while also using their normal development environments, and I believe this pattern unlocks a kind of "AI Middleware". Suddenly, using the sandbox feels exciting, instead of a slog, because it unlocks new avenues of exploration.
Then I discovered the field of AI Control and realized we might just be working on the same problem.
—
Où / Where
UQAM - Pavillon President-Kennedy
PK-R605 (ground floor)
Contact host if you need assistance or for questions