> All the sandbox-exec and all the Lulu rules are not going to protect against something like this.

No, you can block this with sandboxing tools. For my LLM agents, I use a tool called Nono, whose macOS implementation is built on the same foundation as sandbox-exec. One of the key features it supports (and it's not the only tool that does this) is proxying and the injection of credentials via the proxy. For example, when I run Codex at work, Codex does not have access to my Azure AI Foundry API credentials. It sends out requests via dummy credentials, and the proxy injects the real ones outside the sandbox.

The other thing that proxy does is let you restrict network access in various ways. So you can run something like LiteLLM so that it doesn't have access to the internet at large; it can only talk to Azure AI Foundry or AWS Bedrock or whoever your inference provider is. You sandbox LiteLLM separately from your agents, and you give it access to much less. The software firewall stuff isn't perfect but it is a significant defense against things like this.

There are also lots of other restrictions that are easy to do which mitigate this. Nono blocks access to the root of $HOME, then selectively allows access to paths within it, and also explicitly blocks access to common paths to secrets. So it's pretty easy to make sure that access to important things are blocked: SSH credentials, GnuPG socket (and local Unix sockets in general), local OS keyrings, etc.

The way I use tools like this is exclusively by writing policy, testing it, and revising where the policy is inadequate. I never get approval fatigue because I don't do approvals.

You're right that attacks like this are hard to defend against, and that we're used to tool sprawl, and also that LLMs can do things like install tools we don't expect them to themselves! But there are ways to anticipate and mitigate attacks like this. Sandboxing tools (including contained and VMs) that support credential injection are genuinely useful for this.