Hacker News Reader

Comment by skohan | Hacker News Reader

Agent sandboxing/access control is one of the biggest problems to be solved before this technology really should go mainstream.

Even as a technical person, it's not trivial to sandbox agents correctly. The fact that an mis-clicked permission popup could give an agent unrestricted access to a user's disk is a massive risk vector in the hands of lay people who barely understand how any of this works.

So much of current security depends on the model of tying access control to a user account. A lot has to be re-thought in terms of how to grant access to an agent working on the user's behalf, in a way that doesn't make it completely useless, and also doesn't require every user to become a sysadmin managing fine-grained agent permissions manually.

View on HN

I think sandboxing could be solved if effort was put into it. Webassembly seems like a great way to enforce data and execution boundaries for an LLM, for example.

I think the problem is that LLM providers are dis-incentivized from pursuing it because their ethos is gobbling up any and all data they can get.

> Oops, we accidentally yoinked your personal documents, photos, and videos and they’re now swimming in our model’s data ocean! We’re sorrrry, oh well let’s move on.

It’s up to the users to use tools that enforce security/privacy. Open source harnesses like pi.dev seem like a good path forward to me

Sandboxing the agent application is easy. The tough part is sandboxing in such a way that it's still useful.

I.e. if I have an agent running in a WASM sandbox with no access to the host system, I can't ask it to clean up my files. Same thing with things like giving an agent access to your email inbox: doing so allows the agent to provide utility, but it comes with risks, as the agent can delete important emails, or leak sensitive data.

I think a big part of the problem is, a lot of the systems we use and would like agents to help us with don't have a concept of separated roles with different levels of access which can be applied. A lot of times it's all or nothing.

And even when we do have fine-grained access control available, it's a pain in the ass to manage it. Like you can create a GitHub token with fine-grained access control to your repositories and make sure the agent only uses that one to connect, but it's a whole lot easier to use a broad-access token, or just let the agent use your own token, so lots of people will just end up doing that.

And I also like pi, but it's probably one of the worst in terms of sandboxing as it's yolo by default.

Fundamentally i think its how we're trying to skip critically important steps here. Its like the first automobiles that were built with completely uncovered, unfiltered engines. Dropped onto a chassis, and then opened up, dirt, debris, oils, etc intermingled and things blew up. Gas lines, oil reservoirs, compartments and chambers that are all specialized to create a highly efficient and consistent experience came out of designing it properly. The same can be said for Muse and other agents. Without the right scaffolding and harness, of course things go wrong.

I'm a HUGE proponent of putting them behind task gating trees, and sheathing them with QA/QC checks on their processes, especially at this stage. Unfortunately its so easy to create, and all of that takes time and design that many just throw away for getting to results.

Even fine grained ACLs, which are great, dont have the structured approach such autonomous agents need to shore them in (imo).

There is definitely a balance between utility and security/privacy gates. I think people are enjoying the freedom of YOLO for now until the consequences of gate-free agent use become too severe.

> And I also like pi, but it's probably one of the worst in terms of sandboxing as it's yolo by default.

Agreed, but the nice thing about pi is the plugin system and how configurable it is. I can easily hack on the pi harness, whereas a more opinionated one like opencode is more difficult

Yeah I think people are probably widely underestimating the risks. Like if you gave another programmer unlimited access to your system and ssh keys, and they never slept and could code and run terminal commands at dozens or hundreds of words per minute, you would have to trust them a lot to give them that.

And I love pi - it's my daily driver - but the extension system itself is an attack vector. If any process manages to write an extension to your .pi directory, it could rewrite your prompt to have the agent exfiltrate your secrets, or take whatever action on the host system if you don't sandbox it.

> It’s up to the users to use tools that enforce security/privacy.

The problem is that, for most potential users to whom personal (not software dev) agents are being marketed, security & privacy for AI agents is a higher load domain to manage than the things they would want to delegate to agents to relieve load, so taking up the required security & privacy management to avoid problems using the agents creates disutility that exceeds the utility of the agent.

Which is why anyone selling them is going to distract from the issue rather than try to inform their customers.

EDIT:

Another problem is that in many cases, BECAUSE of the assumptions about how people use computers, web services, etc., the required tools DO NOT EXIST. Most account based web services DO NOT have ways to create, say, access tokens with a subset of full account permissions that you could give to an agent rather than full control of the account, because that kind of delegation to limited authority actors was never part of the usage model the vendor designed for.

something tell me that the vast majority of people will give all the permissions the agent ask but even more look for a bypass/yolo permission.

I say that from coding experience. You don’t want to approve 100 windows to get a task done. In the end the only “sane” solution for my setup was a dedicated machine just for the agent with Bitwarden for secrets and full access/yolo mode.

If you are concerned about the agent deleting everything make sure you have a process backing up the git repositories at least to a separate service/hosting and that’s it…for now.

So the solution is to have backups and a way to restore data…

That helps you if the blast radius is on your machine, but it doesn't really help if the agent is using your credentials to cause some damage with some remote system you have access to.

When I was kicking the tires on pi, one of the first things the agent did was push an update to one of my published Rust crates (not the project it was working on).

That in itself wasn't harmful, but it did convince me it was worth the effort to figure out sandboxing after that.

It's already solved. I have two git repos proving these companies can fix the problems. The fact this continues just proves they don't care. In one project I literally containerize CLI coding tools, it works. You might say "sure, but network." I literally wrote a desktop app harness that you can toggle the network on/off too.

This is all amateur hour shenanigans.

I sandbox my coding agents using bubblewrap, but I don't think it's as trivial a problem as you make it sound.

For something like muse that's supposed to be a general-purpose assistant, how do you give it enough access to be useful, without giving it too much access, and creating unacceptable risks? And how do you do that in a way that's comprehensible the average Facebook user who's the target market of this product?

The problem is trivial to solve. Treat AIs like they are users. We have user space for a reason. We have linux namespaces for a reason. We have real airgap architectures where the data center can't even connect out.

When you're a company running around with more than a few billion in the vault, you have no excuse for the level security negligence going on at every phase of rollout.

I gawked at Cursor executing a python script one time and decided enough was enough, I don't raw dog these tools anymore because their developers are the dumbest people to task with security work. They just don't care.

So how would you go about sandboxing Muse in a way that it still functions as a general purpose assistant?

> that it still functions as a general purpose assistant

What features you want it to have? The thing you're saying is super vague, I would just say that one belongs in cloud.

If it must run things on the user's machine, it's gotta be in a rootless container and not be root inside the container. All tools belong in the container. Folder access explicitly configured by the user.

Like I already did this for my local AI: https://github.com/SamInTheShell/loom

It's not perfect or even done, but it works and it shows the security model that should be standard for aligned models we're running.

Unaligned models, absolutely different story.

Also VM is better than container for security, but containers are a bare minimum for me.

Let's say I have an aligned agent, and I want to give it access to my bank account to track my budget, and my email so it can automate responses to certain messages, and automatically unsubscribe from junk emails.

Nether service has a way to configure fine-grained access for a secondary user.

How do I go about giving the agent the ability to perform these tasks without exposing myself to the risk of unexpected destructive behavior from the agent?

Those are API things. Should be an API, build an MCP server for it. Doesn't touch the PC. About the automated responses, that's a write operation, therefore requires user consent.

If you want to design a system specifically to make sure you're not going to get stonewalled for sending me an emdash, you build and maintain a set of permissive rules for sending/reading to avoid my blacklist of people I'll never work with.

Once you got API stuff sorted, just throw a cronjob at it or build a service for that stuff.

^ There be prompts all over the place here, obviously. Realistically your MCP server will end up hacking on POP3.

Its really not as trivial as it sounds, especially as you expand it out into the workforce and you have complex ACLs that are gated by project, scope, budgets, people, etc.

Meta rushed this out, to grab relevancy, especially after losing out to Sam with OpenClaw.

I just dont fundamentally trust such a blackbox with my data. They've literally turned your data and your life into their biggest asset. For it to now have a level of control in your life just seems... irresponsible.

Reply on news.ycombinator.com