[GH-ISSUE #1227] How secure is firejail for unknown programs?

gitea-mirror commented

2026-05-05 06:56:56 -06:00

Owner

Originally created by @peterzhu2118 on GitHub (Apr 16, 2017).
Original GitHub issue: https://github.com/netblue30/firejail/issues/1227

I'm currently writing a program that will run unknown code submitted by other people. Would firejail be a good solution so that if malicious code is submitted, it will not affect the system? Are there cases where outer system resources can be accessed or cases where firejail can be breached?

Originally created by @peterzhu2118 on GitHub (Apr 16, 2017). Original GitHub issue: https://github.com/netblue30/firejail/issues/1227 I'm currently writing a program that will run unknown code submitted by other people. Would firejail be a good solution so that if malicious code is submitted, it will not affect the system? Are there cases where outer system resources can be accessed or cases where firejail can be breached?

gitea-mirror

2026-05-05 06:56:56 -06:00

closed this issue
added the
information_old
label

gitea-mirror commented

2026-05-05 06:57:02 -06:00

Author

Owner

@Ferroin commented on GitHub (Apr 17, 2017):

Whether or not Firejail will work really depends on what degree of security you want and what your specific use case is. Regardless of that, there's always the possibility that there's a bug somewhere that could cause a security breach, but in the case of firejail, it's statistically more likely to be in the kernel than firejail itself (the actually security stuff that firejail does is actually done by the kernel, not firejail itself, it just gets set up by firejail).

Depending on how secure you want things and how much you care about performance, you've got roughly four options:

Run the code in firejail with as much isolation as possible. This will get you reasonable security and probably the best performance short of just running the code un-isolated.
Run the code in a QEMU VM with acceleration enabled. This will get you slightly better isolation than firejail (and present a very different attack surface), but will not get you quite the same performance.
Run the code in a QEMU VM with acceleration disabled (no KVM, no VirtIO). This will provide better isolation than option 2, but will have worse performance.
Run the code in a QEMU VM with acceleration disabled which is itself running inside firejail. Adding firejail outside the QEMU VM should close up any remaining holes, and won't hurt performance much once you're running without acceleration. Unfortunately, you can't use KVM inside a firejail sandbox without the risk of severely compromising the security provided by firejail, so running QEMU accelerated inside firejail isn't exactly an option.

If you're working on something like BOINC (distributed cluster computing), I would suggest looking at what they do, specifically the Leiden Classical project (they let users submit jobs, so I assume they have something in place to provide a security guarantee to their users).

@Ferroin commented on GitHub (Apr 17, 2017): Whether or not Firejail will work really depends on what degree of security you want and what your specific use case is. Regardless of that, there's always the possibility that there's a bug somewhere that could cause a security breach, but in the case of firejail, it's statistically more likely to be in the kernel than firejail itself (the actually security stuff that firejail does is actually done by the kernel, not firejail itself, it just gets set up by firejail). Depending on how secure you want things and how much you care about performance, you've got roughly four options: 1. Run the code in firejail with as much isolation as possible. This will get you reasonable security and probably the best performance short of just running the code un-isolated. 2. Run the code in a QEMU VM with acceleration enabled. This will get you slightly better isolation than firejail (and present a very different attack surface), but will not get you quite the same performance. 3. Run the code in a QEMU VM with acceleration disabled (no KVM, no VirtIO). This will provide better isolation than option 2, but will have worse performance. 4. Run the code in a QEMU VM with acceleration disabled which is itself running inside firejail. Adding firejail outside the QEMU VM should close up any remaining holes, and won't hurt performance much once you're running without acceleration. Unfortunately, you can't use KVM inside a firejail sandbox without the risk of severely compromising the security provided by firejail, so running QEMU accelerated inside firejail isn't exactly an option. If you're working on something like BOINC (distributed cluster computing), I would suggest looking at what they do, specifically the Leiden Classical project (they let users submit jobs, so I assume they have something in place to provide a security guarantee to their users).

gitea-mirror commented

2026-05-05 06:57:04 -06:00

Author

Owner

@peterzhu2118 commented on GitHub (Apr 17, 2017):

@Ferroin Thanks for the great answer! I'm actually looking to write something that allows users to submit their own code to run on my server. For option 4, you are suggesting for me to run a QEMU VM inside of a firejail instance right? Can a QEMU VM reset itself quickly after every submission for something like a "factory reset"? I've used firejail for this purpose before and since it's just a overlay on the kernel, every new instance I create is "resetted" so if one person submitted something malicious, it won't affect the next person.

Do you know any online hosting services (like AWS, Rackspace, Google Cloud, etc.) that provides this kind of sandboxing service?

@peterzhu2118 commented on GitHub (Apr 17, 2017): @Ferroin Thanks for the great answer! I'm actually looking to write something that allows users to submit their own code to run on my server. For option 4, you are suggesting for me to run a QEMU VM inside of a firejail instance right? Can a QEMU VM reset itself quickly after every submission for something like a "factory reset"? I've used firejail for this purpose before and since it's just a overlay on the kernel, every new instance I create is "resetted" so if one person submitted something malicious, it won't affect the next person. Do you know any online hosting services (like AWS, Rackspace, Google Cloud, etc.) that provides this kind of sandboxing service?

gitea-mirror commented

2026-05-05 06:57:05 -06:00

Author

Owner

@Ferroin commented on GitHub (Apr 17, 2017):

Yes, the fourth item was suggesting running a QEMU VM inside a firejail instance (I actually do this on my own systems on occasion when inspecting untrusted code, and adding firejail outside of QEMU imposes almost no overhead). For resetting quickly, most of what you need there is for the actual software in the VM itself to initialize quickly, QEMU can do the equivalent of a hardware reset almost instantly (assuming you have good host hardware of course), and if you go the route of a buildroot based guest, you can have it reset and ready to run the next program in at most a couple of seconds (the best I've managed for something like this in QEMU is 1.3 seconds, but that took a huge amount of work to achieve).

As far as hosting services, AWS EC2 Container service or Amazon Lightsail might do what you need, I'm not certain regarding any other platforms (I've only ever used AWS and Linode for hosting, and Linode probably doesn't provide the degree of security you want). Regardless of that, I would still suggest doing some form of sandboxing inside the hosted instance yourself, just to be extra certain.

@Ferroin commented on GitHub (Apr 17, 2017): Yes, the fourth item was suggesting running a QEMU VM inside a firejail instance (I actually do this on my own systems on occasion when inspecting untrusted code, and adding firejail outside of QEMU imposes almost no overhead). For resetting quickly, most of what you need there is for the actual software in the VM itself to initialize quickly, QEMU can do the equivalent of a hardware reset almost instantly (assuming you have good host hardware of course), and if you go the route of a [buildroot](https://buildroot.org/) based guest, you can have it reset and ready to run the next program in at most a couple of seconds (the best I've managed for something like this in QEMU is 1.3 seconds, but that took a huge amount of work to achieve). As far as hosting services, AWS EC2 Container service or Amazon Lightsail might do what you need, I'm not certain regarding any other platforms (I've only ever used AWS and Linode for hosting, and Linode probably doesn't provide the degree of security you want). Regardless of that, I would still suggest doing some form of sandboxing inside the hosted instance yourself, just to be extra certain.

gitea-mirror commented

2026-05-05 06:57:07 -06:00

Author

Owner

@peterzhu2118 commented on GitHub (Apr 17, 2017):

@Ferroin I have been looking at the AWS EC2 Container service. I haven't had much experience with Docker but from what I can tell it seems to be the perfect solution. Correct me if I'm wrong but each Docker instance seems to be isolated and can be easily reset every time to a default. And since I'm not hosting it myself it should be even more reliable. Am I correct? What can go wrong if I use this service?

@peterzhu2118 commented on GitHub (Apr 17, 2017): @Ferroin I have been looking at the AWS EC2 Container service. I haven't had much experience with Docker but from what I can tell it seems to be the perfect solution. Correct me if I'm wrong but each Docker instance seems to be isolated and can be easily reset every time to a default. And since I'm not hosting it myself it should be even more reliable. Am I correct? What can go wrong if I use this service?

gitea-mirror commented

2026-05-05 06:57:08 -06:00

Author

Owner

@Ferroin commented on GitHub (Apr 17, 2017):

I don't know 100% for certain myself how well Docker isolates things. I do know that both Docker and firejail use the same underlying technologies to isolate things, but they use them differently (for example, I'm pretty sure that firejail doesn't provide any UID remapping, while Docker does). The key difference lies in what they are designed for, firejail is designed to allow a single user to quickly and easily run untrusted code (and from what I've seen, it's one of the best options for this available), while Docker is designed as an application-specific VM or VServer replacement. Given what you've described as your requirements, it sounds like Docker is more in-line with what you want to do than firejail simply because it was designed for this type of thing, while firejail wasn't really.

Now, depending on how much experience you have with coding, it might be best to write your own tool to do this, as Docker may be a bit heavier than you want (not in terms of security, but in terms of runtime and management overhead). If you do go this route, firejail and docker (as well as Google's lmctfy, Gentoo's sandbox program, and a couple of other similar tools) would probably be a good reference on how to do it.

As far as things like failure modes for hosted services, you're probably better off talking with the hosting providers about that.

@Ferroin commented on GitHub (Apr 17, 2017): I don't know 100% for certain myself how well Docker isolates things. I do know that both Docker and firejail use the same underlying technologies to isolate things, but they use them differently (for example, I'm pretty sure that firejail doesn't provide any UID remapping, while Docker does). The key difference lies in what they are designed for, firejail is designed to allow a single user to quickly and easily run untrusted code (and from what I've seen, it's one of the best options for this available), while Docker is designed as an application-specific VM or VServer replacement. Given what you've described as your requirements, it sounds like Docker is more in-line with what you want to do than firejail simply because it was designed for this type of thing, while firejail wasn't really. Now, depending on how much experience you have with coding, it might be best to write your own tool to do this, as Docker may be a bit heavier than you want (not in terms of security, but in terms of runtime and management overhead). If you do go this route, firejail and docker (as well as Google's lmctfy, Gentoo's sandbox program, and a couple of other similar tools) would probably be a good reference on how to do it. As far as things like failure modes for hosted services, you're probably better off talking with the hosting providers about that.

gitea-mirror commented

2026-05-05 06:57:09 -06:00

Author

Owner

@peterzhu2118 commented on GitHub (Apr 19, 2017):

@Ferroin Ok thank you so much for this info! I will try things out to see which one is the best.

@peterzhu2118 commented on GitHub (Apr 19, 2017): @Ferroin Ok thank you so much for this info! I will try things out to see which one is the best.

gitea-mirror referenced this issue

2026-05-05 10:07:33 -06:00

[PR #835] [MERGED] fixed recovery issue #3768

Rows
Columns

[GH-ISSUE #1227] How secure is firejail for unknown programs? #835