tl;dr: I'm building a project to do sandboxing right, I'll be at reInvent, would love to chat there if anyone else is trying to figure this out, too: chase@archodex.com.
People are running around trying to sandbox their agents in creative ways, but I've become jaded by most approaches. Most sandboxes, local or remote, are sandboxing your files and processes away from agents, which is good. But they do hardly anything for network traffic. I don't think file and process isolation alone should count as truly being sandboxed.
At some point agents need to do something, and that means using your credentials to interact with API services. But sandboxes today stop at domain-level filters, meaning they can't tell between read-only, read/write, or control-plane traffic, much less tell a standard GET request for a file on github vs a supply chain attack using github as a c2 channel.
I'm working on a project that tries to solve this. It's still too early to share, but I'm at least dogfooding it myself. It captures the payload of network activity from agent processes and the tools they call, matches it up to the agent's conversation context, parses the salient details from the request, and then allows or blocks traffic based on deterministic and decision-model (i.e. "auto mode") policies. In Linux environments, e.g. Docker Desktop or local/remote sandboxes, it's proxy-less and doesn't disturb the TLS chain, either.
I'm posting today because I'll be at AWS' giant conference in a week and a half. I'll just be hanging around, catching up with friends and past colleagues. But if any of this piques your interest, please hit me up: chase@archodex.com. I'd love to chat with anyone else trying to navigate how to use agents safely.
FWIW, more about the project idea can be seen at https://archodex.com/dex.
Comments URL: https://news.ycombinator.com/item?id=50577102
Points: 1
# Comments: 0

