4 ms·

I keep hearing that `sudo` insecurity argument everywhere. I think that argument is fundamentally wrong. In practice, a package manager that eschews `sudo` doe

I keep hearing that `sudo` insecurity argument everywhere. I think that argument is fundamentally wrong. In practice, a package manager that eschews `sudo` does not protect you any less than a package manager that works with `sudo`. If malware has obtained file system access, it’s already game over. The malicious code can now mess with your shell profile and fake the `sudo` command.

Brian_K_White 6y ago
The main problem isn't really that it doesn't require sudo. If it installed as a user into your home dir, that would be fine. Changing a shared system-wide path or installing shared system-wide files owned by a user is so fundamentally borked that you should immediately reject anything offered by anyone who suggested to do that. But the masses of users who made it popular didn't know that and there were/are a lot of them, and the reason it's wrong won't bite many people because their macbooks only have one user account, and so you have a zillion people who didn't know they were getting bad advice and doing a broken thing, and if you tell them now, they don't see it because they've been doing it wrong for years and their house didn't burn down, and they vastly outnumber the people who unserstand the problem and would never have designed something so broken in the first place, and so it just stays both incorrect and the standard.
- bigbubba 6y ago
  Can you spoonfeed me here? If ~/bin is in my PATH and is writable to my user, what is the problem with /usr/local/bin being the same on my personal computer with a single human user?
  - theamk 6y ago
    Base case (those are true AFAIK): - /usr/local/bin is in a "system-wide" default PATH, (and it is even in front of /bin !) - There is an important service, like Time Machine, which needs root access for important actions -- for example, to erase previous backups - There could be a vulnerability in an application which gets data from the internet -- for the sake of example, let's say it is "foo file viewer". This can be exploited for code execution. --- Case 1: Packages in ~/bin, /usr/local/bin is read-only: A user wants to look at a "foo file" from the internet. The file is malicious, it exploits "foo viewer" and gets local execution. It is a cryptolocker, so it encrypts all the user documents. The malware has installed trojaned "~/bin/sudo" wrapper, but since the user does not use "sudo" that often, it did not have a change to get executed. And none of the system services look into user's "~/bin". A week later, user notices that a document is encrypted. But the time machine backups are still OK. They reformat their machine, and then use time machine backup to restore the documents. Day is saved! --- Case 2: Packages in usr/local/bin: A user wants to look at a "foo file" from the internet. The file is malicious, it exploits "foo viewer" and gets local execution. It is a cryptolocker, so it encrypts all the user documents. The malware has installed a bunch of trojaned binaries, including "/usr/local/bin/touch" binary. Unfortunately, there is a LaunchDaemon which runs the script periodically which contains the line "touch /some/file". That is run as root, so in a short time, the malware gets root access. And it immediately used this access to disable time machine and delete all the backups. A week later, user notices that a document is encrypted. And the time machine backups are gone, too. Oh no! the documents are lost unless they pay the ransom! --- You might notice that some people may consider the scenario unrealistic: What if the user uses command line a lot, and is used to running "sudo"? What if a malware social engineers user by popping up a fake dialog asking for user password? Are there other privilege escalation methods? How often does this happen anyway? You should make your own decisions, but hopefully you at least see what the people are concerned about.
- tsimionescu 6y ago
  Since Unix doesn't (traditionally at least) offer any access control mechanism more granular than a User, it was only natural that you'd get a lot of "system users" that are required for all sorts of operations but are not in any way associated with human users. And now, since you have so many system users and almost always a single human user, you start wanting to install programs that will be used by both yourself and those system users, and it would be extremely problematic to go writing into 10 different bin/ and lib/ folders to install the same thing for 10 users. Not to mention, a lot of software you want to install also runs some daemons or manages some system settings, and obviously requires root access to do so, but mag run into problems if it is not installed for root.
- nemothekid 6y ago
  >Changing a shared system-wide path or installing shared system-wide files owned by a user is so fundamentally borked that you should immediately reject anything offered by anyone who suggested to do that. I don't think that's fair to the homebrew developers. They chose a justifiable tradeoff - as you said most mac systems are single user systems and their design offers better protection to a far larger group of users. When you run `brew install ffmpeg` you are running an arbitrary ruby script from Homebrew git repo and it's probably better that it doesn't run as root. I can't see how that is bad advice; I'd argue that the opposite - sticking to "tradition" despite the changing the landscape of how personal devices have changed security is the broken mindset. I'm sure most macOS users are far more concerned about an arbitrary script exfiltrating their keychain than some contrived multi-user setup where another user installs a keylogger in the system path.
  - Brian_K_White 6y ago
    "contrived" is not an excuse for doing something wrong. All programming, heck all system design of any form, programming, mechanical hardware, legal process, is made of ensuring that situations are handled, including both the obvious ones you thought of and the ones you didn't think of. All "corner cases" could be described as "contrived", yet they happen a million times an hour. "Contrived" is a contrived rationale to ignore something you should not ignore for mere convenience. "contrived" is just a silly and empty accusation, and no argument. If you are in a conversation where the subject deals in probabilities or percentages, then you can talk about liklihood. Having say 1% of customers fail to navigate a subway map is better than having 8% fail. And some contrived example problem like the map needs to be only in shades of green because what if you're colorblind... the contrived nature of the problem matters because the problem deals in percentages where 99% is better than 92%. But in system design (such as application programming or os design, especially the lowest most basic security aspects, things are only possible or not-possible. There is no "likely" and no "contrived". If you can "contrive" an example that invalidates a design, then you have invalidated the design, period. If you cannot contrive an example, then you still have not proven the design, merely you have failed to IN-validate it, yet. The /Applications directory is not equivalent to the /usr/local tree, or any other part of the system. The /Applications dir itself is owned by root, and a user may not change it to be owned by theirself, and a user may only write within it because they are a member of the "admin" group. Until recently, the homebrew installer actually changed the ownership of /usr/local itself (or created it, I don't remember if it maybe didn't exist by default and was created by brew sometimes) rather than just some of the stuff under it. They only changed when forced to, when Apple started more actively protecting system components from modification a couple years ago. If it's ok for shared system components to be owned by joe random user, then why did Apple start protecting /usr/local? For that matter why can't joe random user replace /bin/cp with his own copy, or the kernel? /usr/local/bin/mycp is (or rather should be, and is on sane sysyems) just as much a protected part of the system as /bin/cp. It's only different in that it's locally added to this machine and didn't come on the install cd, and may have different behavior than other parts of the system expect etc. But it's not "steve's" either. A "mycp" owned by "steve" should only exist in /home/steve (or /Users/steve on mac). If brew installed everything in /Applications/Brew.app that would be ok too, because then it would be just doing what other apps do, and because /Applications is a special case which has extra handling by other means than merely the traditional unix file ownerships. That would still be mildly broken because it would mean that all users could use the brew binaries and other assets, but only the installing user could modify them. That is fine for most apps which do not modify their internal contents, but brew is essentially a whole system itself, and it's contents are modified during normal use. A theoretical Brew.app is not like Calculator.app So, the only sane way to do it is to install as the user somewhere within the users own domain (their home dir) or install in a system-wide shared place and require sudo for any "brew update" or "make install". If it were all owned by a special user and/or group "brew" (not whatever random user happened to do the install), that would be sane too, but not necessary.
- pram 6y ago
  That is exactly how installing an application in MacOS always works. When you run an installer for a native app, and it goes into /Applications, it's owned by the user who installed it. What Homebrew is doing isn't functionally different, other than /usr/local is in your PATH I guess?
segfaultbuserr 6y ago
> fake the `sudo` command. I personally see that "su" and "sudo" are both fundamentally insecure. Compromising any sudoer, and root is essentially lost (e.g. keyloggers), thus the "wheel" group should only contain trusted accounts. It can be more secure if all root operations require a separate console, isolated from the rest of the world. For example, on historical Unix servers, it was a common practice to restrict root logins to a physical terminal. But nowadays it's unfeasible. One solution for servers is disabling "su" and "sudo", and to allow root logins via pubkeys. This goes against the best practices, but can be more secure when it's used properly. Alternatively, create an "admin" account with sudo privilege, and guard it as carefully as your root account, because a sudoer is root. On desktop, a solution can be using Intel ME/SGX/TXT and allowing it to take over your screen and keyboard to create a root shell - but it is proprietary, unauditable, and cannot be trusted. But it's not strictly needed - pure software can do something similar, e.g. you run two X servers on two separate VTs with two users, only the second user has access to a root shell. I guess better integration at the graphics server/system level is possible, you just need to be careful about GUI spoofing. For example, on Windows NT, pressing Control+Alt+Delete opens an unspoofable & isolated secure login screen - perhaps we should implement a secure root shell like that! A more practical, although much less secure alternative is requiring 2FA, e.g. pressing a physical button on a USB token to sudo, it's something you can use today. But it cannot solve the problem of malicious command injections. But on second thought, these problems are not really a priority in infosec - user applications are insecure in general, and user data is accessible without root anyway... As xkcd said, "they can read my mails, take my money, and impersonate me, but at least they cannot install device drivers without my permission..." So doing all of the works can be a total waste of time. Also, even with a separate root console, one must also have strong discipline not to execute anything or opening any complex binary data from another user (potentially compromised), which makes the whole idea fragile. Finally QubesOS's approach is an innovative workaround and worth mentioning - normal and sensitive applications are running in separate VMs, providing strong isolation. Since both system and user applications are already isolated, it's not really that dangerous to be "root", so by default, everyone can sudo without any authentication. But if you are concerned about someone attacking supervisor via root, you can use Dom0 to 2FA authenticate sudo in an application VM without any extra hardware.
- 6y ago
IshKebab 6y ago
Yes the only way around that is to have sudo activate a Secure Attention Key. Windows has had that for many many years ("press ctrl-alt-del to continue") though they seem to have ditched it in Windows 10. Presumably users didn't know what it meant or why it was there and just did it when asked by anyone. I don't think it's really possible to make sudo use a SAK on Linux though because it requires tight integration of lots of components and Linux systems are cobbled together with glue and string.
- __turbobrew__ 6y ago
  The magic sysrq key combination works, I don’t see why you couldn’t do something similar for a SAK?
- theamk 6y ago
  Are you saying you want "sudo" to use a a password entry mechanism which cannot be spoofed or intercepted? Something very similar to this is pretty easy on Linux. If you are willing to pay a little, requiring Yuibikey press for sudo is supported [0]. If you don't want extra hardware, you can use "pam_exec" PAM method to read second factor (like PIN) from "/dev/input/by-id/...". This'll read from specific input device directly, bypassing GUI session or any other processes controlled by non-root user. Or use a "chvt" call instead, and you can have a secure display output too. But while this will prevent "evil sudo" from capturing your password, it won't prevent "evil sudo" modifying the commands you want executed (the same problem exists with Windows SAK / Escalation prompt: you know you are approving some action, but it is not clear which action exactly). I guess Linux has a unique advantage in that "sudo" is designed for user-readable commands -- so in many cases, confirming the command on un-tamperable output will be enough for safety (for example, "sudo apt install foo" will be safe this way). But you'd still want to make sure you are vigilant at all times -- if you have a habit of typing "curl ... | sudo sh", then no type of SAK process will save you from evil sudo. [0] https://support.yubico.com/hc/en-us/articles/360016649099-Ubuntu-Linux-Login-Guide-U2F https://support.yubico.com/hc/en-us/articles/360016649099-Ub...
  - IshKebab 6y ago
    A SAK is some action that only the kernel can detect. On Windows user-level programs cannot receive the ctrl-alt-del key sequence, so if you do that and something happens you can be sure you are interacting with the real authentication GUI, not a fake one. I don't think any of your solutions provide that capability.

Brian_K_White 6y ago

The main problem isn't really that it doesn't require sudo. If it installed as a user into your home dir, that would be fine. Changing a shared system-wide path or installing shared system-wide files owned by a user is so fundamentally borked that you should immediately reject anything offered by anyone who suggested to do that. But the masses of users who made it popular didn't know that and there were/are a lot of them, and the reason it's wrong won't bite many people because their macbooks only have one user account, and so you have a zillion people who didn't know they were getting bad advice and doing a broken thing, and if you tell them now, they don't see it because they've been doing it wrong for years and their house didn't burn down, and they vastly outnumber the people who unserstand the problem and would never have designed something so broken in the first place, and so it just stays both incorrect and the standard.

bigbubba 6y ago
Can you spoonfeed me here? If ~/bin is in my PATH and is writable to my user, what is the problem with /usr/local/bin being the same on my personal computer with a single human user?
- theamk 6y ago
  Base case (those are true AFAIK): - /usr/local/bin is in a "system-wide" default PATH, (and it is even in front of /bin !) - There is an important service, like Time Machine, which needs root access for important actions -- for example, to erase previous backups - There could be a vulnerability in an application which gets data from the internet -- for the sake of example, let's say it is "foo file viewer". This can be exploited for code execution. --- Case 1: Packages in ~/bin, /usr/local/bin is read-only: A user wants to look at a "foo file" from the internet. The file is malicious, it exploits "foo viewer" and gets local execution. It is a cryptolocker, so it encrypts all the user documents. The malware has installed trojaned "~/bin/sudo" wrapper, but since the user does not use "sudo" that often, it did not have a change to get executed. And none of the system services look into user's "~/bin". A week later, user notices that a document is encrypted. But the time machine backups are still OK. They reformat their machine, and then use time machine backup to restore the documents. Day is saved! --- Case 2: Packages in usr/local/bin: A user wants to look at a "foo file" from the internet. The file is malicious, it exploits "foo viewer" and gets local execution. It is a cryptolocker, so it encrypts all the user documents. The malware has installed a bunch of trojaned binaries, including "/usr/local/bin/touch" binary. Unfortunately, there is a LaunchDaemon which runs the script periodically which contains the line "touch /some/file". That is run as root, so in a short time, the malware gets root access. And it immediately used this access to disable time machine and delete all the backups. A week later, user notices that a document is encrypted. And the time machine backups are gone, too. Oh no! the documents are lost unless they pay the ransom! --- You might notice that some people may consider the scenario unrealistic: What if the user uses command line a lot, and is used to running "sudo"? What if a malware social engineers user by popping up a fake dialog asking for user password? Are there other privilege escalation methods? How often does this happen anyway? You should make your own decisions, but hopefully you at least see what the people are concerned about.
tsimionescu 6y ago
Since Unix doesn't (traditionally at least) offer any access control mechanism more granular than a User, it was only natural that you'd get a lot of "system users" that are required for all sorts of operations but are not in any way associated with human users. And now, since you have so many system users and almost always a single human user, you start wanting to install programs that will be used by both yourself and those system users, and it would be extremely problematic to go writing into 10 different bin/ and lib/ folders to install the same thing for 10 users. Not to mention, a lot of software you want to install also runs some daemons or manages some system settings, and obviously requires root access to do so, but mag run into problems if it is not installed for root.
nemothekid 6y ago
>Changing a shared system-wide path or installing shared system-wide files owned by a user is so fundamentally borked that you should immediately reject anything offered by anyone who suggested to do that. I don't think that's fair to the homebrew developers. They chose a justifiable tradeoff - as you said most mac systems are single user systems and their design offers better protection to a far larger group of users. When you run `brew install ffmpeg` you are running an arbitrary ruby script from Homebrew git repo and it's probably better that it doesn't run as root. I can't see how that is bad advice; I'd argue that the opposite - sticking to "tradition" despite the changing the landscape of how personal devices have changed security is the broken mindset. I'm sure most macOS users are far more concerned about an arbitrary script exfiltrating their keychain than some contrived multi-user setup where another user installs a keylogger in the system path.
- Brian_K_White 6y ago
  "contrived" is not an excuse for doing something wrong. All programming, heck all system design of any form, programming, mechanical hardware, legal process, is made of ensuring that situations are handled, including both the obvious ones you thought of and the ones you didn't think of. All "corner cases" could be described as "contrived", yet they happen a million times an hour. "Contrived" is a contrived rationale to ignore something you should not ignore for mere convenience. "contrived" is just a silly and empty accusation, and no argument. If you are in a conversation where the subject deals in probabilities or percentages, then you can talk about liklihood. Having say 1% of customers fail to navigate a subway map is better than having 8% fail. And some contrived example problem like the map needs to be only in shades of green because what if you're colorblind... the contrived nature of the problem matters because the problem deals in percentages where 99% is better than 92%. But in system design (such as application programming or os design, especially the lowest most basic security aspects, things are only possible or not-possible. There is no "likely" and no "contrived". If you can "contrive" an example that invalidates a design, then you have invalidated the design, period. If you cannot contrive an example, then you still have not proven the design, merely you have failed to IN-validate it, yet. The /Applications directory is not equivalent to the /usr/local tree, or any other part of the system. The /Applications dir itself is owned by root, and a user may not change it to be owned by theirself, and a user may only write within it because they are a member of the "admin" group. Until recently, the homebrew installer actually changed the ownership of /usr/local itself (or created it, I don't remember if it maybe didn't exist by default and was created by brew sometimes) rather than just some of the stuff under it. They only changed when forced to, when Apple started more actively protecting system components from modification a couple years ago. If it's ok for shared system components to be owned by joe random user, then why did Apple start protecting /usr/local? For that matter why can't joe random user replace /bin/cp with his own copy, or the kernel? /usr/local/bin/mycp is (or rather should be, and is on sane sysyems) just as much a protected part of the system as /bin/cp. It's only different in that it's locally added to this machine and didn't come on the install cd, and may have different behavior than other parts of the system expect etc. But it's not "steve's" either. A "mycp" owned by "steve" should only exist in /home/steve (or /Users/steve on mac). If brew installed everything in /Applications/Brew.app that would be ok too, because then it would be just doing what other apps do, and because /Applications is a special case which has extra handling by other means than merely the traditional unix file ownerships. That would still be mildly broken because it would mean that all users could use the brew binaries and other assets, but only the installing user could modify them. That is fine for most apps which do not modify their internal contents, but brew is essentially a whole system itself, and it's contents are modified during normal use. A theoretical Brew.app is not like Calculator.app So, the only sane way to do it is to install as the user somewhere within the users own domain (their home dir) or install in a system-wide shared place and require sudo for any "brew update" or "make install". If it were all owned by a special user and/or group "brew" (not whatever random user happened to do the install), that would be sane too, but not necessary.
pram 6y ago
That is exactly how installing an application in MacOS always works. When you run an installer for a native app, and it goes into /Applications, it's owned by the user who installed it. What Homebrew is doing isn't functionally different, other than /usr/local is in your PATH I guess?

segfaultbuserr 6y ago

> fake the `sudo` command. I personally see that "su" and "sudo" are both fundamentally insecure. Compromising any sudoer, and root is essentially lost (e.g. keyloggers), thus the "wheel" group should only contain trusted accounts. It can be more secure if all root operations require a separate console, isolated from the rest of the world. For example, on historical Unix servers, it was a common practice to restrict root logins to a physical terminal. But nowadays it's unfeasible. One solution for servers is disabling "su" and "sudo", and to allow root logins via pubkeys. This goes against the best practices, but can be more secure when it's used properly. Alternatively, create an "admin" account with sudo privilege, and guard it as carefully as your root account, because a sudoer is root. On desktop, a solution can be using Intel ME/SGX/TXT and allowing it to take over your screen and keyboard to create a root shell - but it is proprietary, unauditable, and cannot be trusted. But it's not strictly needed - pure software can do something similar, e.g. you run two X servers on two separate VTs with two users, only the second user has access to a root shell. I guess better integration at the graphics server/system level is possible, you just need to be careful about GUI spoofing. For example, on Windows NT, pressing Control+Alt+Delete opens an unspoofable & isolated secure login screen - perhaps we should implement a secure root shell like that! A more practical, although much less secure alternative is requiring 2FA, e.g. pressing a physical button on a USB token to sudo, it's something you can use today. But it cannot solve the problem of malicious command injections. But on second thought, these problems are not really a priority in infosec - user applications are insecure in general, and user data is accessible without root anyway... As xkcd said, "they can read my mails, take my money, and impersonate me, but at least they cannot install device drivers without my permission..." So doing all of the works can be a total waste of time. Also, even with a separate root console, one must also have strong discipline not to execute anything or opening any complex binary data from another user (potentially compromised), which makes the whole idea fragile. Finally QubesOS's approach is an innovative workaround and worth mentioning - normal and sensitive applications are running in separate VMs, providing strong isolation. Since both system and user applications are already isolated, it's not really that dangerous to be "root", so by default, everyone can sudo without any authentication. But if you are concerned about someone attacking supervisor via root, you can use Dom0 to 2FA authenticate sudo in an application VM without any extra hardware.

6y ago

IshKebab 6y ago

Yes the only way around that is to have sudo activate a Secure Attention Key. Windows has had that for many many years ("press ctrl-alt-del to continue") though they seem to have ditched it in Windows 10. Presumably users didn't know what it meant or why it was there and just did it when asked by anyone. I don't think it's really possible to make sudo use a SAK on Linux though because it requires tight integration of lots of components and Linux systems are cobbled together with glue and string.

__turbobrew__ 6y ago
The magic sysrq key combination works, I don’t see why you couldn’t do something similar for a SAK?
theamk 6y ago
Are you saying you want "sudo" to use a a password entry mechanism which cannot be spoofed or intercepted? Something very similar to this is pretty easy on Linux. If you are willing to pay a little, requiring Yuibikey press for sudo is supported [0]. If you don't want extra hardware, you can use "pam_exec" PAM method to read second factor (like PIN) from "/dev/input/by-id/...". This'll read from specific input device directly, bypassing GUI session or any other processes controlled by non-root user. Or use a "chvt" call instead, and you can have a secure display output too. But while this will prevent "evil sudo" from capturing your password, it won't prevent "evil sudo" modifying the commands you want executed (the same problem exists with Windows SAK / Escalation prompt: you know you are approving some action, but it is not clear which action exactly). I guess Linux has a unique advantage in that "sudo" is designed for user-readable commands -- so in many cases, confirming the command on un-tamperable output will be enough for safety (for example, "sudo apt install foo" will be safe this way). But you'd still want to make sure you are vigilant at all times -- if you have a habit of typing "curl ... | sudo sh", then no type of SAK process will save you from evil sudo. [0] https://support.yubico.com/hc/en-us/articles/360016649099-Ubuntu-Linux-Login-Guide-U2F https://support.yubico.com/hc/en-us/articles/360016649099-Ub...
- IshKebab 6y ago
  A SAK is some action that only the kernel can detect. On Windows user-level programs cannot receive the ctrl-alt-del key sequence, so if you do that and something happens you can be sure you are interacting with the real authentication GUI, not a fake one. I don't think any of your solutions provide that capability.