The first 3 “hardening” points are not great.

Essentially it’s just: remove .py files an execute del os.environ[“SESSION_TOKEN“]? This doesn’t really sound very secure, there are a number of ways to bypass both of these.

It’s just security through obscurity

The billion engineers building sandbox tools at the moment are missing the point. Sandboxing doesn't matter when the LLM is vulnerable to prompt injection. Every MCP server you install, every webpage it fetches, every file it reads is a threat. Yeah you can sit there and manually approve every action it takes, but then how is any of this useful when you have to supervise it constantly? Even Anthropic say that this doesn't work because reviewing every action leads to exhaustion and rubber stamping.

The problem is not what the LLM shouldn't have access to, it's what it does have access to.

The usefulness of LLMs is severely limited while they lack the ability to separate instructions and data, or as Yann LeCun said, predict the consequences of their actions.

I think this is pretty standard and similar to approaches that are evolving naturally (I've certainly used very similar patterns).

I'd be pretty keen to actually hear more about the Unikraft setup and other deeper details about the agent sandboxes regarding the tradeoffs and optimizations made. All the components are there but has someone open-sourced a more plug-and-play setup like this?

It’s neat to see more projects adopting Unikernals. I’ve played around with Unikraft’s Cloud offering about a year ago when it was CLI/API only and was impressed by the performance but found too many DX and polish issues to take it to production. Looks like they’ve improved a lot of that since.

maybe the usecase that makes unikernels alot more mainstream. Always found them intriguing