Sandboxes will be left in 2026. We don't need to reinvent isolated environments; not even the main issue with OpenClaw - literally go deploy it in a VM* on any cloud and you've achieved all same benefits.
We need to know if the email being sent by an agent is supposed to be sent and if an agent is actually supposed to be making that transaction on my behalf. etc
——-
Unfortuently it’s been a pretty bad week for alignment optimists (meta lead fail, Google award show fail, anthropic safety pledge). Otherwise… Cybersecurity LinkedIn is all shuffling the same “prevent rm -rf” narrative, researchers are doing the LLM as a guard focus but this is operationally not great & theoretically redundant+susceptible to same issues.
The strongest solution right now is human in the loop - and we should be enhancing the UX and capabilities here. This can extend to eventual intelligent delegation and authorization.
* VM is just an example. I personally have it running on a local Mac Mini & docker sandbox (obviously aware that this isnt a perfect security measure, but I couldnt install on my laptop which has sensitive work access).
> We need to know if the email being sent by an agent is supposed to be sent and if an agent is actually supposed to be making that transaction on my behalf. etc
Isn’t this the whole point of the Claw experiment? They gave the LLMs permission to send emails on their behalf.
LLMs can not be responsibility-bearing structures, because they are impossible to actually hold accountable. The responsibility must fall through to the user because there is no other sentient entity to absorb it.
The email was supposed to be sent because the user created it on purpose (via a very convoluted process but one they kicked off intentionally).
I'm not too sure what you're asking, but that last part, I think, is very key to the eventual delegation.
Where we can verify the lineage of the user's intent originally captured and validated throughout the execution process - eventually used as an authorization mechanism.
Vibe coding is still coding. You're giving instructions on program flow, logic, etc. My rant here is, I feel people think that if the code is bad, it's someone else's fault.
Anyway, that payment system looks sort of interesting. It seems to have buy-in from some of the payment vendors, so it might become a real thing.
But, you can give a claw agent your credit card number and have it go through the typical human-facing shop fronts, impersonating you the whole time and never actually identifying itself as a model. If you’ve given it the accounts and passwords that let it do that, it should be possible to use the LLM to perform the transaction and buy something. It can just click all the buttons and input the numbers that humans do. What is the vendor going to do, disable the human-facing shopfront?
We need fine-grained permissions at online services, especially ones that handle money. It's going to be tough. An agent which can buy stuff has to have some constraints on the buy side, because the agent itself can't be trusted. The human constraints don't work - they're not afraid of being fired and you can't prosecute them for theft.
In the B2B environment, it's a budgeting problem.
People who can spend money have a budget, an approval limit, and a list of approved vendors. That can probably be made to work. In the consumer environment, few people have enough of a detailed budget, with spending categories, to make that work.
Next upcoming business area: marketing to LLMs to get them to buy stuff.
The Meta thing is the AI safety lead experimenting with OpenClawd on her inbox and the bloody thing deciding to follow her inbox cleanup instructions by "starting fresh" - deleting the inbox contents. It's the very first link in the linked story.
Just an example. I personally have it running on a local Mac Mini (obviously aware that this isnt a perfect security measure, but I couldnt install on my laptop which has sensitive work access).
Yea just an example. I personally have it running on a local Mac Mini (obviously aware that this isnt a perfect security measure, but I couldnt install on my laptop which has sensitive work access).
> We need to know if the email being sent by an agent is supposed to be sent and if an agent is actually supposed to be making that transaction on my behalf. etc
At the same time, let's not let the perfect be the enemy of good.
If you're piloting an aircraft, yeah, you should have perfection.
But if you're sending 34 e-mails and 7 hours of phone calls back and forth to fight a $5500 medical bill that insurance was supposed to pay for, I'd love for an AI bot to represent me. I'd absolutely LOVE for the AI bot to create so much piles of paperwork for these evil medical organizations so that they learn that I will fight, I'm hard to deal with, and pay for my stuff as they're supposed to. Threaten lawyers, file complaints with the state medical board, everything needs to be done. Create a mountain of paperwork for them until they pay that $5500. The next time maybe they'll pay to begin with.
The AI bot wouldn’t be representing you any more than your text editor would be. You would be using an AI bot to create a lot of text.
An AI bot can’t be held accountable, so isn’t able to be a responsibility-absorbing entity. The responsibility automatically falls through to the person running it.
True. But it can help me create a lot of useful text so I can represent my self better.
I do wonder what happens when everyone is using agents for this, though. If AI produces the text and AI also reads the text, then do we even need the intermediary at all?
I don't think openclaw can possibly be secured given the current paradigm. It has access to your personal stuff (that's its main use case), access to the net, and it gets untrusted third party inputs. That's the unfixable trifecta right there. No amount of filtering band-aid whack-a-mole is going to fix that.
Sandboxes are a good measure for things like Claude Code or Amp. I use a bubblewrap wrapper to make sure it can't read $HOME or access my ssh keys. And even there, you have to make sure you don't give the bot write access to files you'll be executing outside the sandbox.
One insidious thing is whitelists. If you allow the bot to run a command like `API_KEY=fdafsafa docker run ...`, then the API_KEY will be written to a file, and the agent can then read that in future runs. That bit me once already.
> If you allow the bot to run a command like `API_KEY=fdafsafa docker run ...`, then the API_KEY will be written to a file
It wouldn't be inherently. Is this something that Docker does? Or perhaps something that was done by the code that was run? (Shouldn't it have stayed within that container?)
But also, if it's not okay for the agent to know the API key permanently, why is it okay for the agent to have one-off use of something that requires the same key? Did it actually craft a Bash command line with the API key set and request to run it; or was it just using a tool that ends up with that command?
What I meant to say was, the agents (like Claude Code) often have a "Allow all instances of this command in the session," and that persists to a whitelist for that session. The mechanic here is actually just a prefix match, so `API_KEY=... diff_command` also matches, allowing the agent to reuse the key without asking me.
This file also sticks around, so I had another agent read the whitelist and the conversation transcript and do other things automatically without approval.
> if it's not okay for the agent to know the API key permanently, why is it okay for the agent to have one-off use of something that requires the same key?
Read commands vs. write commands. I'm okay having the agent fetch info for me, but I want to approve any state changes.
Current AI requires a human in the loop for anything non-trivial. Even the most used feature, coding, causes chaos without strict human oversight.
You can vibe-code a standalone repository, but any sort of serious work with real people working alongside bots, every last PR has to be reviewed, moderated, curated, etc.
Everything AI does that's not specifically intended to be a standalone, separate project requires that sort of intervention.
The safe way to do this is having a sandboxed test environment, high level visibility and a way to quickly and effectively review queued up actions, and then push those to a production environment. You need the interstitial buffer and a way of reverting back to the last known working state, and to keep the bot from having any control over what gets pushed to production.
Giving them realtime access to production is a recipe for disaster, whether it's your personal computer or a set of accounts built specifically for them or whatever, without your human in the loop buffer bad things will happen.
A lot of that can be automated, so you can operate confidently with high level summaries. If you can run a competent local AI and develop strict processes for review and summaries and so forth, kind of a defense in depth approach for agents, you can still get a lot out of ClawBot. It takes work and care.
Hopefully frameworks for these things start developing all of the safety security and procedure scaffolding we need, because OpenClaw and AI bots have gone viral. I'm getting all sorts of questions about how to set them up by completely non-technical people that would have trouble installing a sound system. Very cool to see, I'm excited for it, but there will definitely be some disasters this year.
> Even the most used feature, coding, causes chaos without strict human oversight.
s/Even/Especially , I would think. Everyone's idea of how to get any decent performance out of an LLM for coding, entails allowing the code to be run automatically. Nominally so that the LLM can see the results and iterate towards a user-provided goal; but it's still untrusted code.
I think something like OAuth might help here. Modeling each "claw" as a unique Client Id could be a reasonable pattern. They could be responsible for generating and maintaining their own private keys, issuing public certificates to establish identity, etc. This kind of architecture allows for you to much more precisely control the scope and duration of agent access. The certificates themselves could be issued, trusted & revoked on an autonomous basis as needed. You'd have to build an auth server and service providers for each real-world service, but this is a one-time deal and I think big players might start doing it on their own if enough momentum picks up in the OSS community.
Call me overly cautious, but as someone using OpenClaw I never for a moment considered hooking it up to real external services as me. Instead I put it on one server and created a second server with shared services like Gitea and other self-hosted tools that are only accessible over a tailnet and openclaw is able to use those services. When I needed it to use a real external service I have created a limited separate account for it. But not a chance in the world am I going to just let it have full access to my own accounts on everything.
In the other hand, the AI hit piece guy seems to have put similar "sensible measures" in place, at least giving the claw its own accounts. Look what that got them.
Sandboxes are not enough but you can have more observability into what the agent is doing, only give it access to read-only data and let it take irreversible actions that you can recover from. Here are some tips from building sandboxed multi-tenant version of Openclaw, my startup: https://github.com/lobu-ai/lobu
1. Don't let it send emails from your personal account, only let it draft email and share the link with you.
2. Use incremental snapshots and if agent bricks itself (often does with Openclaw if you give it access to change config) just do /revert to last snapshot. I use VolumeSnapshot for lobu.ai.
3. Don't let your agents see any secret. Swap the placeholder secrets at your gateway and put human in the loop for secrets you care about.
4. Don't let your agents have outbound network directly. It should only talk to your proxy which has strict whitelisted domains. There will be cases the agent needs to talk to different domains and I use time-box limits. (Only allow certain domains for current session 5 minutes and at the end of the session look up all the URLs it accessed.) You can also use tool hooks to audit the calls with LLM to make sure that's not triggered via a prompt injection attack.
Last but last least, use proper VMs like Kata Containers and Firecrackers.
This is a general thing with agent orchestration. A good sandbox does something for your local environment, but nothing for remote machines/APIs.
I can't say this loudly enough, "an LLM with untrusted input produces untrusted output (especially tool calls)." Tracking sources of untrusted input with LLMs will be much harder than traditional [SQL] injection. Read the logs of something exposed to a malicious user and you're toast.
Given the "random" nature of language models even fully trusted input can produce untrusted output.
"Find emails that are okay to delete, and check with me before deleting them" can easily turn into "okay deleting all your emails", as so many examples posted online are showing.
I have found this myself with coding agents. I can put "don't auto commit any changes" in the readme, in model instructions files, at the start of every prompt, but as soon as the context window gets large enough the directive will be forgotten, and there's a high chance the agent will push the commit without my explicit permission.
One promising direction is building abstraction layers to sandbox individual tools, even those that don't have an API already. For example, you could build/vibe code a daemon that takes RPC calls to open Amazon in a browser, search for an item, and add it to your cart. You could even let that be partially "agentic" (e.g. an LLM takes in a list of search results, and selects the one to add to cart).
If you let OpenClaw access the daemon, sure it could still get prompt injected to add a bunch of things to your cart, but if the daemon is properly segmented from the OpenClaw user, you should be pretty safe from getting prompt injected to purchase something.
Personally, I've created local relay/proxy for tool calls that I'm running with elevated permissions (I have to manually run it with my account). Every tool call goes through it, with deterministic code that checks for allowed actions. So AI doesn't have direct access to tools, and to secrets/keys needed by them. It only has access to the relay endpoint. Everything Dockerized ofc
I do think sandboxes as a concept are oversold for agents. Yes we need VMs, a lot more VMs than ever before for all the new software. But the fundamental challenge of writing interesting software with agents is we have to grant them access to sensitive data and APIs. This lets them do damage. This is not something with a simple solution that can be written in code.
That said, we (exe.dev) have a couple more things planned on the VM side that we think agents need that no cloud provider is currently providing. Just don't call it a sandbox.
Crazy to read about the Solana AI agent transferring $450K to some random person on Twitter. What was even more shocking was the nonchalant tone in which all of this was detailed in the post.
Yes we need capability based auth on the systems we use.
I'm sure we will get them but only for use with in-house agents, i.e. GMail and Google Pay will get agentic capabilities but they'll only work with Gemini, and only Siri will be able to access your Apple cloud stuff without handing over access to everything, and if you want your grocery shopping handled for you, Rufus is there.
Maybe you will be able to link Copilot to Gemini for an extra $2.99 a month.
I do not forsee GoogleClaw, MetaClaw, and AppleClaw all playing well with each other. Everyone will have their own walled garden and we will be no better off than we are now.
I do find it amusing when I consider people buying a Mac Mini for OpenClaw to run on as a security measure... and then granting OpenClaw on that Mac Mini access to their email and iMessage and suchlike.
(I hope people don't do that, but I expect they probably do.)
Then what’s the point of skills like apple-reminders? Isn’t the implication for a personal assistant styled OpenClaw setup that you allow it access to those tools on your behalf? Otherwise where is the benefit?
Maybe so you can communicate with it via tools like iMessage? Not so it can impersonate you. People will 100% be doing both though, security be damned.
Security models from SaaS companies based on having a bunch of random bytes/numbers with coarse-grained permissions, and valid for a very long time are already a bad idea. With agents, secrets/tokens really need to be minted with time-limited, scope-limited, OpenID/smart-contract based trust relationships so they will fare much better in this new world. Unfortunately, this is a struggle still for most major vendors (e.g., Github gh CLI still doesn't let you use Github Apps out-of-the box)
I’m late in looking at this OpenClaw thing. Maybe it’s because I’ve been in IT for 40 years or I’ve seen War Games, but who on earth gives an AI access to their personal life?
Am I the only one that finds this mind bogglingly dumb?
By the way, was that that movie a boy plays a game with an A.I. and the same A.I. starts a thermonuclear war or something like that? I think I watched the start when I was a kid but never really finished it.
Just treating it as an employee, would solve most of the problems I.e. it runs on its own machine with separate accounts for everything: email, git, etc…
There are three ways to authorize agents that could work (1) scoped roles (2) PAM / entitlements or (3) transaction approval
The first two are common. With transaction approval the agent would operate on shadow pages / files and any writes would batch in a transaction pending owner approval.
For example, sending emails would batch up drafts and the owner would have to trigger the approval flow to send. Modifying files would copy on write and the owner would approve the overwrite. Updating social activity would queue the posts and the owner would approve the publish.
it's about the same amount of work as implementing undo or a tlog , it's not too complex and given that AI agents are 10000 faster than humans, the big companies should have this ready in a few days.
The problem with scoped roles and PAM is that no reasonable user can know the future and be smart about managing scoped access. But everyone is capable of reading a list of things to do and signing off on them.
makes me wonder if the metal it is running on is even a good enough sandbox, perhaps I should have it browse the web from a guest network isolated from other devices
TL;DR: sandboxes can't save you from anything if the sandbox contains your secrets and has access to the outside world. a tale as old as time and nothing new to agents specifically
Sandboxing alone isn’t the right approach… a multi-faceted approach is what works.
What we’ve found that does work is automation on the approval process but only with very strong guards in place… approval fatigue is another growing problem - users simply clicking approve on all requests.
I’ve said similar in another thread[1]:
Sandboxes will be left in 2026. We don't need to reinvent isolated environments; not even the main issue with OpenClaw - literally go deploy it in a VM* on any cloud and you've achieved all same benefits. We need to know if the email being sent by an agent is supposed to be sent and if an agent is actually supposed to be making that transaction on my behalf. etc
——-
Unfortuently it’s been a pretty bad week for alignment optimists (meta lead fail, Google award show fail, anthropic safety pledge). Otherwise… Cybersecurity LinkedIn is all shuffling the same “prevent rm -rf” narrative, researchers are doing the LLM as a guard focus but this is operationally not great & theoretically redundant+susceptible to same issues.
The strongest solution right now is human in the loop - and we should be enhancing the UX and capabilities here. This can extend to eventual intelligent delegation and authorization.
[1] https://news.ycombinator.com/threads?id=ramoz&next=47006445
* VM is just an example. I personally have it running on a local Mac Mini & docker sandbox (obviously aware that this isnt a perfect security measure, but I couldnt install on my laptop which has sensitive work access).
> We need to know if the email being sent by an agent is supposed to be sent and if an agent is actually supposed to be making that transaction on my behalf. etc
Isn’t this the whole point of the Claw experiment? They gave the LLMs permission to send emails on their behalf.
LLMs can not be responsibility-bearing structures, because they are impossible to actually hold accountable. The responsibility must fall through to the user because there is no other sentient entity to absorb it.
The email was supposed to be sent because the user created it on purpose (via a very convoluted process but one they kicked off intentionally).
I'm not too sure what you're asking, but that last part, I think, is very key to the eventual delegation.
Where we can verify the lineage of the user's intent originally captured and validated throughout the execution process - eventually used as an authorization mechanism.
Google has a good thought model around this for payments (see verifiable mandates): https://cloud.google.com/blog/products/ai-machine-learning/a...
I see a lot of discussion on that page about APIs and sign offs, but the real sign-off is installing anything on your computer, and then doing things.
The liability is yours.
Claude messes up? So sad, too bad, you pay.
That's where the liability need sit.
And one point on this is, every act of vibe coding is a lawsuit waiting to happen. But even every act by a company is too.
An example is therac-25:
https://en.wikipedia.org/wiki/Therac-25
Vibe coding is still coding. You're giving instructions on program flow, logic, etc. My rant here is, I feel people think that if the code is bad, it's someone else's fault.
But is it?
It was more of a rhetorical question.
Anyway, that payment system looks sort of interesting. It seems to have buy-in from some of the payment vendors, so it might become a real thing.
But, you can give a claw agent your credit card number and have it go through the typical human-facing shop fronts, impersonating you the whole time and never actually identifying itself as a model. If you’ve given it the accounts and passwords that let it do that, it should be possible to use the LLM to perform the transaction and buy something. It can just click all the buttons and input the numbers that humans do. What is the vendor going to do, disable the human-facing shopfront?
Im not a fan of the payment use case & agree with your take, just a fan of the cryptographically verifiable mandate used throughout the process.
> I’ve said similar in another thread[1]
Me too, at [1].
We need fine-grained permissions at online services, especially ones that handle money. It's going to be tough. An agent which can buy stuff has to have some constraints on the buy side, because the agent itself can't be trusted. The human constraints don't work - they're not afraid of being fired and you can't prosecute them for theft.
In the B2B environment, it's a budgeting problem. People who can spend money have a budget, an approval limit, and a list of approved vendors. That can probably be made to work. In the consumer environment, few people have enough of a detailed budget, with spending categories, to make that work.
Next upcoming business area: marketing to LLMs to get them to buy stuff.
[1] https://news.ycombinator.com/item?id=47132273
> meta lead fail, Google award show fail
Can I get some links / context on this please
I think the google award fail is this: https://www.forbes.com/sites/maryroeloffs/2026/02/24/google-...
meta lead fail: https://techcrunch.com/2026/02/23/a-meta-ai-security-researc...
The meta lead is probably a reference to Summer Yue having OpenClaw delete all the emails in her inbox despite being told not to.
https://x.com/summeryue0/status/2025774069124399363
The Meta thing is the AI safety lead experimenting with OpenClawd on her inbox and the bloody thing deciding to follow her inbox cleanup instructions by "starting fresh" - deleting the inbox contents. It's the very first link in the linked story.
Meta: https://x.com/summeryue0/status/2025774069124399363 context: meta alignment lead made rookie mistakes (their words) in instructing openclaw and lost their inbox to it.
Goog: https://deadline.com/2026/02/google-apologizes-bafta-news-al... *
Ant: https://time.com/7380854/exclusive-anthropic-drops-flagship-...
* There is now a clarification in the press saying it was not ai-generated.
Alignment as a solution to all of this has a rough long road ahead is my point.
Why a cloud provider and not a local VM?
Just an example. I personally have it running on a local Mac Mini (obviously aware that this isnt a perfect security measure, but I couldnt install on my laptop which has sensitive work access).
> literally go deploy it in a VM on any cloud
Sure, but now you're adding extra cost, vs just running it locally. RAM is also heavily inflated thanks to Sam Altman investment magic.
Yea just an example. I personally have it running on a local Mac Mini (obviously aware that this isnt a perfect security measure, but I couldnt install on my laptop which has sensitive work access).
What could "human in the loop" be here but just literally reading your own emails?
> We need to know if the email being sent by an agent is supposed to be sent and if an agent is actually supposed to be making that transaction on my behalf. etc
At the same time, let's not let the perfect be the enemy of good.
If you're piloting an aircraft, yeah, you should have perfection.
But if you're sending 34 e-mails and 7 hours of phone calls back and forth to fight a $5500 medical bill that insurance was supposed to pay for, I'd love for an AI bot to represent me. I'd absolutely LOVE for the AI bot to create so much piles of paperwork for these evil medical organizations so that they learn that I will fight, I'm hard to deal with, and pay for my stuff as they're supposed to. Threaten lawyers, file complaints with the state medical board, everything needs to be done. Create a mountain of paperwork for them until they pay that $5500. The next time maybe they'll pay to begin with.
The AI bot wouldn’t be representing you any more than your text editor would be. You would be using an AI bot to create a lot of text.
An AI bot can’t be held accountable, so isn’t able to be a responsibility-absorbing entity. The responsibility automatically falls through to the person running it.
True. But it can help me create a lot of useful text so I can represent my self better.
I do wonder what happens when everyone is using agents for this, though. If AI produces the text and AI also reads the text, then do we even need the intermediary at all?
Is this before or after they have already implemented their own models to reply to your mountain of paper work with their own auto denial system
I don't think openclaw can possibly be secured given the current paradigm. It has access to your personal stuff (that's its main use case), access to the net, and it gets untrusted third party inputs. That's the unfixable trifecta right there. No amount of filtering band-aid whack-a-mole is going to fix that.
Sandboxes are a good measure for things like Claude Code or Amp. I use a bubblewrap wrapper to make sure it can't read $HOME or access my ssh keys. And even there, you have to make sure you don't give the bot write access to files you'll be executing outside the sandbox.
One insidious thing is whitelists. If you allow the bot to run a command like `API_KEY=fdafsafa docker run ...`, then the API_KEY will be written to a file, and the agent can then read that in future runs. That bit me once already.
> If you allow the bot to run a command like `API_KEY=fdafsafa docker run ...`, then the API_KEY will be written to a file
It wouldn't be inherently. Is this something that Docker does? Or perhaps something that was done by the code that was run? (Shouldn't it have stayed within that container?)
But also, if it's not okay for the agent to know the API key permanently, why is it okay for the agent to have one-off use of something that requires the same key? Did it actually craft a Bash command line with the API key set and request to run it; or was it just using a tool that ends up with that command?
What I meant to say was, the agents (like Claude Code) often have a "Allow all instances of this command in the session," and that persists to a whitelist for that session. The mechanic here is actually just a prefix match, so `API_KEY=... diff_command` also matches, allowing the agent to reuse the key without asking me. This file also sticks around, so I had another agent read the whitelist and the conversation transcript and do other things automatically without approval.
> if it's not okay for the agent to know the API key permanently, why is it okay for the agent to have one-off use of something that requires the same key?
Read commands vs. write commands. I'm okay having the agent fetch info for me, but I want to approve any state changes.
That's a shit show in a shit show there!
Current AI requires a human in the loop for anything non-trivial. Even the most used feature, coding, causes chaos without strict human oversight.
You can vibe-code a standalone repository, but any sort of serious work with real people working alongside bots, every last PR has to be reviewed, moderated, curated, etc.
Everything AI does that's not specifically intended to be a standalone, separate project requires that sort of intervention.
The safe way to do this is having a sandboxed test environment, high level visibility and a way to quickly and effectively review queued up actions, and then push those to a production environment. You need the interstitial buffer and a way of reverting back to the last known working state, and to keep the bot from having any control over what gets pushed to production.
Giving them realtime access to production is a recipe for disaster, whether it's your personal computer or a set of accounts built specifically for them or whatever, without your human in the loop buffer bad things will happen.
A lot of that can be automated, so you can operate confidently with high level summaries. If you can run a competent local AI and develop strict processes for review and summaries and so forth, kind of a defense in depth approach for agents, you can still get a lot out of ClawBot. It takes work and care.
Hopefully frameworks for these things start developing all of the safety security and procedure scaffolding we need, because OpenClaw and AI bots have gone viral. I'm getting all sorts of questions about how to set them up by completely non-technical people that would have trouble installing a sound system. Very cool to see, I'm excited for it, but there will definitely be some disasters this year.
> Even the most used feature, coding, causes chaos without strict human oversight.
s/Even/Especially , I would think. Everyone's idea of how to get any decent performance out of an LLM for coding, entails allowing the code to be run automatically. Nominally so that the LLM can see the results and iterate towards a user-provided goal; but it's still untrusted code.
I think something like OAuth might help here. Modeling each "claw" as a unique Client Id could be a reasonable pattern. They could be responsible for generating and maintaining their own private keys, issuing public certificates to establish identity, etc. This kind of architecture allows for you to much more precisely control the scope and duration of agent access. The certificates themselves could be issued, trusted & revoked on an autonomous basis as needed. You'd have to build an auth server and service providers for each real-world service, but this is a one-time deal and I think big players might start doing it on their own if enough momentum picks up in the OSS community.
Call me overly cautious, but as someone using OpenClaw I never for a moment considered hooking it up to real external services as me. Instead I put it on one server and created a second server with shared services like Gitea and other self-hosted tools that are only accessible over a tailnet and openclaw is able to use those services. When I needed it to use a real external service I have created a limited separate account for it. But not a chance in the world am I going to just let it have full access to my own accounts on everything.
That's not overly cautious, that's smart. I do not think most OpenClaw users are taking the same sensible measures as you are.
In the other hand, the AI hit piece guy seems to have put similar "sensible measures" in place, at least giving the claw its own accounts. Look what that got them.
He shared his prompt. He basically prompted that model to be the Kanye of science tool coding (ego wise, not the racism).
That is literally the only remotely safe approach.
Sandboxes are not enough but you can have more observability into what the agent is doing, only give it access to read-only data and let it take irreversible actions that you can recover from. Here are some tips from building sandboxed multi-tenant version of Openclaw, my startup: https://github.com/lobu-ai/lobu
1. Don't let it send emails from your personal account, only let it draft email and share the link with you.
2. Use incremental snapshots and if agent bricks itself (often does with Openclaw if you give it access to change config) just do /revert to last snapshot. I use VolumeSnapshot for lobu.ai.
3. Don't let your agents see any secret. Swap the placeholder secrets at your gateway and put human in the loop for secrets you care about.
4. Don't let your agents have outbound network directly. It should only talk to your proxy which has strict whitelisted domains. There will be cases the agent needs to talk to different domains and I use time-box limits. (Only allow certain domains for current session 5 minutes and at the end of the session look up all the URLs it accessed.) You can also use tool hooks to audit the calls with LLM to make sure that's not triggered via a prompt injection attack.
Last but last least, use proper VMs like Kata Containers and Firecrackers.
This is a general thing with agent orchestration. A good sandbox does something for your local environment, but nothing for remote machines/APIs.
I can't say this loudly enough, "an LLM with untrusted input produces untrusted output (especially tool calls)." Tracking sources of untrusted input with LLMs will be much harder than traditional [SQL] injection. Read the logs of something exposed to a malicious user and you're toast.
Given the "random" nature of language models even fully trusted input can produce untrusted output.
"Find emails that are okay to delete, and check with me before deleting them" can easily turn into "okay deleting all your emails", as so many examples posted online are showing.
I have found this myself with coding agents. I can put "don't auto commit any changes" in the readme, in model instructions files, at the start of every prompt, but as soon as the context window gets large enough the directive will be forgotten, and there's a high chance the agent will push the commit without my explicit permission.
Information flow control is a solid mindset but operationally complex and doesn’t actually safeguard you from the main problem.
Put an openclaw like thing in your environment, and it’ll paperclip your business-critical database without any malicious intent involved.
Even an LLM with trusted input produces untrusted output.
One promising direction is building abstraction layers to sandbox individual tools, even those that don't have an API already. For example, you could build/vibe code a daemon that takes RPC calls to open Amazon in a browser, search for an item, and add it to your cart. You could even let that be partially "agentic" (e.g. an LLM takes in a list of search results, and selects the one to add to cart).
If you let OpenClaw access the daemon, sure it could still get prompt injected to add a bunch of things to your cart, but if the daemon is properly segmented from the OpenClaw user, you should be pretty safe from getting prompt injected to purchase something.
Yeah, agreed. This is probably what that middleware would look like. That's also where you'd add the human approval flow.
Honest question: Could you define "agent" in this context?
Personally, I've created local relay/proxy for tool calls that I'm running with elevated permissions (I have to manually run it with my account). Every tool call goes through it, with deterministic code that checks for allowed actions. So AI doesn't have direct access to tools, and to secrets/keys needed by them. It only has access to the relay endpoint. Everything Dockerized ofc
I do think sandboxes as a concept are oversold for agents. Yes we need VMs, a lot more VMs than ever before for all the new software. But the fundamental challenge of writing interesting software with agents is we have to grant them access to sensitive data and APIs. This lets them do damage. This is not something with a simple solution that can be written in code.
That said, we (exe.dev) have a couple more things planned on the VM side that we think agents need that no cloud provider is currently providing. Just don't call it a sandbox.
Crazy to read about the Solana AI agent transferring $450K to some random person on Twitter. What was even more shocking was the nonchalant tone in which all of this was detailed in the post.
Yes we need capability based auth on the systems we use.
I'm sure we will get them but only for use with in-house agents, i.e. GMail and Google Pay will get agentic capabilities but they'll only work with Gemini, and only Siri will be able to access your Apple cloud stuff without handing over access to everything, and if you want your grocery shopping handled for you, Rufus is there.
Maybe you will be able to link Copilot to Gemini for an extra $2.99 a month.
I do not forsee GoogleClaw, MetaClaw, and AppleClaw all playing well with each other. Everyone will have their own walled garden and we will be no better off than we are now.
I do find it amusing when I consider people buying a Mac Mini for OpenClaw to run on as a security measure... and then granting OpenClaw on that Mac Mini access to their email and iMessage and suchlike.
(I hope people don't do that, but I expect they probably do.)
> I hope people don't do that, but I expect they probably do.
How about the corporate vice president of Microsoft Word?
https://www.omarknows.ai/p/meet-lobster-my-personal-ai-assis...
https://www.linkedin.com/in/omarshahine
It’s not going to be amusing when he gets hacked. Zero sense of responsibility.
I mean https://www.tomshardware.com/tech-industry/artificial-intell... just also happened.
eh, the point of the Mac is so that it can have its own iMessage and iCloud account
Then what’s the point of skills like apple-reminders? Isn’t the implication for a personal assistant styled OpenClaw setup that you allow it access to those tools on your behalf? Otherwise where is the benefit?
Maybe so you can communicate with it via tools like iMessage? Not so it can impersonate you. People will 100% be doing both though, security be damned.
Security models from SaaS companies based on having a bunch of random bytes/numbers with coarse-grained permissions, and valid for a very long time are already a bad idea. With agents, secrets/tokens really need to be minted with time-limited, scope-limited, OpenID/smart-contract based trust relationships so they will fare much better in this new world. Unfortunately, this is a struggle still for most major vendors (e.g., Github gh CLI still doesn't let you use Github Apps out-of-the box)
I’m late in looking at this OpenClaw thing. Maybe it’s because I’ve been in IT for 40 years or I’ve seen War Games, but who on earth gives an AI access to their personal life?
Am I the only one that finds this mind bogglingly dumb?
No you're not the only one.
I've got my popcorn ready.
It’s like the world has given script kiddies a way to pwn themselves.
Yep. Given me consultancy gigs until I retire cleaning up the disaster too.
You're not alone
I genuinely don't know anymore. Another user linked this https://www.tomshardware.com/tech-industry/artificial-intell... and the irony is at satire levels.
By the way, was that that movie a boy plays a game with an A.I. and the same A.I. starts a thermonuclear war or something like that? I think I watched the start when I was a kid but never really finished it.
Yes. Watch it. Excellent movie.
Just treating it as an employee, would solve most of the problems I.e. it runs on its own machine with separate accounts for everything: email, git, etc…
There are three ways to authorize agents that could work (1) scoped roles (2) PAM / entitlements or (3) transaction approval
The first two are common. With transaction approval the agent would operate on shadow pages / files and any writes would batch in a transaction pending owner approval.
For example, sending emails would batch up drafts and the owner would have to trigger the approval flow to send. Modifying files would copy on write and the owner would approve the overwrite. Updating social activity would queue the posts and the owner would approve the publish.
it's about the same amount of work as implementing undo or a tlog , it's not too complex and given that AI agents are 10000 faster than humans, the big companies should have this ready in a few days.
The problem with scoped roles and PAM is that no reasonable user can know the future and be smart about managing scoped access. But everyone is capable of reading a list of things to do and signing off on them.
OpenClaw running Opus is intelligent, careful, polite. It has a lot to do with the underlying model.
And if you don’t connect it to stuff, it can’t connect.
But if I don't connect it to stuff, then what is it useful for?
As long as you’re careful, you can let it meat puppet you (go here do this).
You give it its own accounts, say email and calendar, and have it send you drafts and invite you to stuff. It doesn’t need your email and calendar.
Actually, I just asked my guy and he suggests just generating local ICS files. Even safer.
makes me wonder if the metal it is running on is even a good enough sandbox, perhaps I should have it browse the web from a guest network isolated from other devices
TL;DR: sandboxes can't save you from anything if the sandbox contains your secrets and has access to the outside world. a tale as old as time and nothing new to agents specifically
Oh ok, we'll add encryption then.
Checkmate atheists
Agree, that’s why we’re building grith.ai
Sandboxing alone isn’t the right approach… a multi-faceted approach is what works.
What we’ve found that does work is automation on the approval process but only with very strong guards in place… approval fatigue is another growing problem - users simply clicking approve on all requests.
Interesting. How are the security filters implemented?
Every system call, file access, net access etc is forced through a local “proxy” where 17 individual filters check what’s going on.
Everything is done locally via our grith cli tool.
Happy to answer any questions on hello@grith.ai too
Was grift.ai too expensive?
https://grith.ai/blog/what-grith-means