Remix.run Logo
▲ zaptheimpaler 19 hours ago

They're totally right to not bother with security inside a harness, its just a terrible approach that creates permission fatigue and bad UX. Claude asking me to approve a 50 line arcane bash command doesn't help at all. Pi also has the same philosophy of not enforcing security within the harness.

Security should be enforced outside the harness with a sandbox. You can define what's allowed and what's not in one place and have strong enforcement guarantees instead of ad-hoc permission requests every minute.

▲neobrain 16 hours ago | parent | next [-]

> Security should be enforced outside the harness with a sandbox. You can define what's allowed and what's not in one place and have strong enforcement guarantees instead of ad-hoc permission requests every minute.

I agree in principle, but in practice most sandboxing technology (like bwrap) is too static for this use case: You grant access once during startup and can't extend/revoke without a restart. That's inconvenient when you realize halfway through your session that you need additional permissions, and it's insufficient for permissions that are only briefly needed (for the duration of one tool call).

A harness that controls the sandbox (e.g. for each individual tool invocation) can manage this much more effectively.

▲jeremyjh 13 hours ago | parent | next [-]

I almost never have to extend permissions of my sandbox. By default it can read/write the project tree, and read most files on the machine with specific secret locations excluded. If I do have to expand permissions and restart, I just resume the session.

▲s1ugh34d_ 16 hours ago | parent | prev | next [-]

A harness that controls the sandbox is exactly what I’m after. Tools don’t always need to share the same space, and I can isolate work much better if separate toolchains are isolated in sandboxes. My biggest gripe with the major harnesses is how they sandbox tool calls.

▲or1gaminal 11 hours ago | parent | prev [-]

Offtop: "I agree in principle, but in practice..." do you listen to The Rest is History podcast? I just heard this line for the first time in this episode https://www.youtube.com/watch?v=8QIGseLdLk0 and now it props up second time on the same day!

▲neobrain 4 hours ago | parent [-]

Nope. Frequency illusion in action :)

▲rpunkfu 18 hours ago | parent | prev | next [-]

Security when calling external services is much more than just FS permissions, sandbox doesn’t protect at all what you push to github or call API with

▲ghosty141 18 hours ago | parent | next [-]

Doesn't change his argument that sandboxing should be done outside the harness. You can also handle network access etc externally.

▲testtestmetest 14 hours ago | parent [-]

You really wanting handle this at the sanbox layer? I have 100 different containers configurations, only a few differnet harness configurations.

You're just being stubborn, the best UX is through the harness. Most users dont even know how to configure a sandbox properly and dont even use one.

Ive known quite a few naive people that have gotten their home directories wiped and really gotten burned. Many users of these products dont even know how to use a terminal and these harness companies are marketing to the average joe.

▲krzyk 13 hours ago | parent | next [-]

> Most users dont even know how to configure a sandbox properly and dont even use one.

This is their problem. But running and screaming "ai deleted my database" is all on them, not on LLM.

▲ghosty141 14 hours ago | parent | prev | next [-]

I have a sandbox script and a simple wrapper that runs any command in that sandbox thus that "pi" -> without sandbox, "pisb" -> pi in the sandbox. I practically only use pisb.

This separation doesn't mean it must result in bad UX. It just needs somebody to make a nice UI/TUI to configure the sandbox and when running that it does its setup and open whatever command you want.

The problem with the harness doing sandboxing is that from my experience the simply break all the time leading to users ignoring properly configuring it and using abominations like codex' `Approve for me` where IT ASKS A SEPARATE AI and not the user.

▲testtestmetest 13 hours ago | parent [-]

My main issue is these are being marketed to average joes. You wont see anything on these harnesses documentation, opencode et al, who say it should be used in a sandbox. The people making them, opencode et al, scream everyday about how you're holding yourself back by using permissions without a single mention of sandbox.

Hundreds of thousands of people are using them naively, if they shipped with common sense strictish permissions instead of none at all, like what opencode opted to do it would be safer for everyone. Its much easier for a normal non programmer to configure a simple config file than it is for them to learn how to properly setup a container/vm, they will just not do the latter 9/10.

Honestly, a harness should always install heavily gaurded, remember when Grok code/build was instantly indexing your home directory and sending it back to HQ on the first request it made..

▲ghosty141 13 hours ago | parent [-]

Like I wrote above I generally agree that the sandboxing should by enabled "by default" and should not be an extra opt-in that you have to know about first.

That being said I still believe this should be separated and composable at an application level and power users should be able to switch out the sandbox. This way you don't have to trust the harness since it doesn't know anything about the sandbox and can't modify it etc.

▲whoanoway 13 hours ago | parent | prev [-]

[dead]

▲trollbridge 13 hours ago | parent | prev [-]

You can make restricted GitHub accounts and API tokens.

▲CyberShadow 14 hours ago | parent | prev | next [-]

OpenCode's security problems are not limited to just permissioning agent actions. For a long time, simply starting OpenCode, and without requiring any other actions or prior configuration, allowed any website to run any code on your computer without any restrictions.

▲DanielHB 17 hours ago | parent | prev | next [-]

Yeah I came to the same conclusion, I migrated to run harnesses and my projects inside a container. However some kinds of development are really hard to run inside a container (like native mobile development or 3d game engines).

I have even been running my IDE in server/client mode with the server inside the container too (because of malicious extensions).

▲trollbridge 13 hours ago | parent | prev | next [-]

100%.

I have my agent request permissions for things like a Cloudflare API key (or I guide it or just come up with the request myself). These go in a simple queue, with justifications.

I review them and then have another agent with far more permissions create the tokens, with expiry and so on, and then hand the API tokens back.

The original agent doesn’t even run as my main user; it sits on a computer in a restricted account (easy to do with either macOS or Linux, or even Windows these days).

It has taken some rather long prompts to get it to self-report on itself when it tries to workaround.

▲testtestmetest 14 hours ago | parent | prev | next [-]

I dont want it doing damage to the sanbox either. I have half a dozen sandboxes running 24/7, there's definately a balance. My sandbox can't say dont run python for this type of task or similar conditional permissions. There are gaurdrails that the harness should handle well.

Also the harness should definately just have a solid permission structure for the average user that 1000% is not using a sandbox, most people I know naively do not and think these things are perfectly safe.

▲californical 17 hours ago | parent | prev | next [-]

Even better, just limit what the agent can access. Don’t give it bash, give it the exact collection of tools you want it to be able to use. Limit it, it still works totally fine without full bash access. I gave my harness access to everything it needs through tool calls to specifically scoped actions and it works great

▲hsaliak 16 hours ago | parent | prev | next [-]

this is ultimately the right approach. You are going to get more mileage by giving your agent a specific set of tools and operations it can freely access and restrict the rest. Ive also found that any dynamic rejections pollute the context and may lead the agent to go find worse workarounds.

Put your agent in a restricted but fixed environment, and let it do its task. Anything else is mostly theatre and token burn.

▲hendersoon 15 hours ago | parent | prev [-]

Wrong. This is largely a solved problem with “auto review” permissions mode, which uses another LLM to review all tool calls that could potentially be dangerous. That solution isn’t perfect of course but it is reliable and usable at the point. This is available in Claude code, codex, copilot, and a bunch of others.

I find auto review mode mandatory and refuse to consider any harness that doesn’t support it.

Opencode doesn’t support it. And there are no good plugins to do it either, the most popular had only six stars last I checked.

▲edgyquant 11 hours ago | parent | next [-]

Auto review often refuses to do things I directly ask the LLm to do, the other day it refused to go to a personal blog and copy the contents for a prototype so I could test a new design. Have had it refuse to use test data on numerous occasions and only tell me when I manually checked and realized it wouldn’t do it

▲csh0 12 hours ago | parent | prev [-]

No so sure about it being “largely a solved problem”

https://embracethered.com/blog/posts/2026/breaking-claude-co...