From: Jean-loup Gailly [jloup@gailly.net] Sent: Thursday, September 26, 2002 4:57 PM To: bugtraq@securityfocus.com Subject: remote SYSTEM compromise in WASD OpenVMS http server -----BEGIN PGP SIGNED MESSAGE----- Hash: SHA1 Multiple vulnerabilities in WASD http server for OpenVMS Version 1.0, 25 Sept 2002. 0. Contents 1. Summary 2. Severity: Critical 3. Vulnerable versions 4. Description 5. Solutions 6. Examples of site weaknesses 7. Conclusion 8. Acknowledgments 9. Document history 1. Summary WASD VMS Hypertext Services is a popular http server for OpenVMS released under the GNU GPL. See http://wasd.vsm.com.au/WASD/ The default installation of the WASD server allows: - universal directory traversal - instant access to the entire web server tree - trivial bypassing of access control rules - getting the location of the document root - read access to the whole web server configuration - read access to all web server logs - disclosure of directories supposed to be hidden - getting the list of all cgi scripts - getting the sources of all cgi scripts - read access to OpenVMS system files - user home directories might be readable - one very serious flaw in a cgi script enabled by default - some problems with other cgi-scripts enabled by default 2. Severity: Critical When combining different vulnerabilities, a remote SYSTEM (root) compromise is possible in a default installation of WASD versions up to 8.0. Even without compromising the system, important files which are supposed to stay confidential can easily be read remotely. 3. Vulnerable versions WASD 7.1, 7.2 (up to 7.2.3), 8.0, and possibly earlier versions. WASD 8.1, and the update versions 8.0.1 and 7.2.4, will fix the known vulnerabilities. 4. Description The main problems are: - the default configuration is much too liberal - the access control rules can be trivially bypassed - cgi scripts are run by default under the identity of the main server There are also problems with some cgi scripts provided by default, at least one of which is very serious. 4.1 The default configuration is much too liberal 4.1.1 The builtin "tree" script By default the entire directory tree of the web server can be seen with http://webserver/tree/ Documentation: http://wasd.vsm.com.au/ht_root/doc/env/env_0400.html#43 Example: http://wasd.vsm.com.au/tree/ 4.1.2 "/*.*" directory traversal Universal directory traversal is builtin with http://webserver/dirname/*.* even if there is a page /dirname/index.html Documentation: http://wasd.vsm.com.au/ht_root/doc/env/env_0400.html 4.1.3 The builtin "upd" script A nice graphical interface for directory traversal is builtin with http://webserver/upd/dirname/ even if there is a page /dirname/index.html Documentation: http://wasd.vsm.com.au/httpd/-/updhelp.html 4.1.4 The builtin search All documents can be searched by default. Documentation: http://wasd.vsm.com.au/ht_root/doc/env/env_0700.html#97 4.1.5 Most of the server root is accessible by default The document root is the main server root, which includes everything (configuration files, scripts, executables, etc...). Directory traversal is enabled on most of the directories and access to the other directories is possible anyway (see later). The WASD documentation states that the liberal option was on purpose: http://wasd.vsm.com.au/ht_root/doc/htd/htd_0600.html#98 "The default configuration is fairly liberal, providing information of use in a technical environment, but that may be superfluous or less-than-desirable in other, possibly commercial environments." OpenVMS has a reputation of being a very secure operating system. This may explain why VMS system administrators may have a false sense of security when installing the WASD web server and leave the default configuration almost untouched. The default configuration in WASD 8.1 will be much more restrictive. 4.2 The access control rules can be trivially bypassed All the liberal features described above can be used to bypass the few restrictions set by the web server. 4.2.1 Bypassing access restrictions set by the web server The configuration file httpd$map.conf contains rules such as: pass /ht_root/wwwroot* fail /ht_root/* fail /-/* In this example, the server root is ht_root and the document root is wwwroot. The "fail" rules attempt to restrict access to the server root but can be trivially bypassed with: http://theserver/ht_root/wwwroot/-/*.* (note: "-" is the VMS equivalent of ".."). So the only real protection comes from the directory protections and ACLs (access control lists) imposed by the operating system. (See section "Solutions" below.) 4.2.2 The location of the document root can easily be obtained The location of the document root can easily be obtained with the "where" builtin script. Documentation: http://wasd.vsm.com.au/ht_root/doc/htd/htd_2100.html#364 Even if the "where" script is not enabled, the "404 not found" message gives away the document root. On a server where the document root was (correctly) not the entire web server root (logical name HT_ROOT), the real document root could still be obtained with http://theserver/notfound which gives "Document not found ... /ht_root/wwwroot/notfound" This allows getting the entire web server root with http://theserver/ht_root/wwwroot/-/*.* as described above. The full physical path may also be given hidden as a comment in the html returned in the "404 not found" message: 4.2.3 The full web server configuration can easily be obtained The web server configuration file in ht_root/local/httpd$map.conf is generally supposed to be protected by a fail rule: fail /ht_root/local/* However this can often be trivially bypassed as shown above: http://theserver/ht_root/wwwroot/-/local/httpd$map.conf On one of the machines tested, there was even no "fail" rule so the configuration could be obtained directly as http://theserver/local/httpd$map.conf The configuration file httpd$map.conf gives a lot of information to intruders, in particular all the access control rules, all the script directories, all the virtual domains handled, all the accessible directories not under ht_root, etc... 4.2.4 All the web server logs can easily be obtained This is just a variation on the above. There is generally a rule fail /ht_root/log/* but all the logs can generally be obtained with http://theserver/upd/ht_root/src/-/log/ unless they are protected with adequate ACLs. If the logs are protected by ACLs but you have a user account on the machine hosting the web server, then some of the logs may still be obtained because the last request may be available in the logical name (environment variable) HTTPD80$REQUEST : $ show log HTTPD80$REQUEST "HTTPD80$REQUEST" = "08 11:56:42.200.430.5287.0.9000.http://theserver: 80.ip.address.of.caller.GET /filename" The logical name is dynamically updated at each request. However this was observed on only one system with an old version of WASD so the problem may be fixed in 8.0. 4.2.5 The "tree" builtin script shows directories supposed to be hidden The server logs are actually in a subdirectory /ht_root/log/server/ Since there is generally a fail rule for /ht_root/log/* the subdirectory server is supposed to be hidden, but the "tree" builtin script happily shows it: http://theserver/tree/ht_root/ 4.2.6 Directory protection can be bypassed anyway Some directories are better protected and are not even visible with the "tree" builtin scripts. On one tested site, the directory /ht_root/script_local/ exists but cannot be seen with /ht_root/script_local/*.* or with /tree/. This site has probably used the following configuration described in http://wasd.vsm.com.au/ht_root/doc/htd/htd_0600.html#99 [DirAccess] - Make "disabled" to completely remove the ability to generate directory listings under any circumstances. However the search feature comes to rescue, and the directory (together will all its scripts, see next section) can still be seen with http://theserver/.../*.com?search=$ If even the standard search feature has been disabled, one can also try the default glist script, which is supposed to give only a list of images but also gives all subdirectories: http://theserver/cgi-bin/glist/ht_root/?list=now In any case, on this particular server the directory script_local is given in /robots.txt so it is not difficult to see that it exists. 4.2.7 The list of cgi scripts can easily be obtained It is not possible to see the list of cgi scripts with the url http://theserver/cgi-bin/*.* however the scripts are generally in ht_root/script/ or ht_root/script_local. If a configuration rule is supposed to block access to these directories (either because the document root is correctly separated from the server root, or because there is an explicit rule blocking script/*), it is generally possible to get the list of scripts and their sources with an url such as: http://theserver/ht_root/wwwroot/-/script_local/*.* For cgi scripts of users, which can be run with url http://theserver/~username/cgi-bin/scriptname it is a bit less trivial to obtain the list of scripts because the obvious attempts fail: http://theserver/~username/cgi-bin/*.* but the following generally works: http://theserver/~username/xxx/-/cgi-bin/*.* where xxx doesn't have to be an existing directory. If it doesn't work, then the search feature can generally be used: http://theserver/~username/.../*.com?search=$ (note: on OpenVMS "..." indicates directory recursion). This searches all the *.com scripts for the character $ which always starts a command on OpenVMS, and so gives as a result the full list of all .com scripts under all subdirectories of the home directory. It is even possible to search *.* instead of *.com. 4.2.8 The sources of cgi scripts can easily be obtained After the list of scripts has been obtained as above, it is trivial to get the sources of the scripts by clicking on a link. For general server scripts the link can be of the form: http://theserver/ht_root/wwwroot/-/script_local/scriptname For user scripts obtained by the search feature, the link is: http://theserver/extract/~username/cgi-bin/scriptname.com?highlight=$ Here "extract" is a script provided by default by WASD. It allows getting the source of a script instead of executing it, even if the directory containing the script is named cgi-bin. 4.2.9 OpenVMS system files can generally be read On several tested sites, a configuration rule is supposed to give access to only a selected portion of system files: pass /sys$common/syslib/* /sys$common/syslib/* However other system files can easily be read, for example http://theserver/sys$common/syslib/-/sysmgr/systartup_vms.com 4.2.10 User home directories might be readable It is common practice to map url http://theserver/~username/ to a subdirectory /user_disk/username/web/ of the home directory /user_disk/username/. But if the OpenVMS protections and ACLs on the home directory are not set correctly, it is possible to traverse it with: http://theserver/~username/-/*.* Actually this often returns an error because of a strange mapping rule: pass /*/-/* /ht_root/runtime/*/* but it is easy to work around this rule with: http://theserver/~username/x/--/*.* where x doesn't have to be the name of an existing directory, and "--" represents for VMS the equivalent of "../.." In one of the sites tested, there was even no mapping to a subdirectory so the whole user home directory was available with http://theserver/~username/*.* 4.3 CGI scripts are run by default under the identity of the main server By default the server runs the image httpd.exe (or httpd_ssl.exe) under the identity of user http$server. CGI scripts are also run by default under the same identity. Thus a flaw in one CGI script can affect the entire server. On Unix, the main server is typically run as root but CGI scripts are typically run under user "nobody". Thus a bad script cannot affect the entire server. The WASD server allows running CGI scripts as a user other than http$server. This will be the default in WASD 8.1. 4.4 Problems with some cgi scripts enabled by default I have not studied all scripts provided in a default installation of WASD. But at least one is very dangerous and there are some bugs in others. 4.4.1 Write to an arbitrary file on the web server One script enabled by default allows writing contents of the attacker's choice to an arbitrary file on the server, as long as the VMS ACLs allows it. This flaw can be exploited to get a remote SYSTEM (root) compromise. Given the severity of this flaw, no details are given here. See the "Solutions" section below for temporary workarounds. 4.4.2 Severe leakage of information in cgi_process.com Running http://theserver/cgi-bin/cgi_process gives a *lot* of useful information for an intruder. In particular it gives all the privileges owned by the script while running. (For Unix users: a script run with privileges is somewhat equivalent to a setuid or setgid program, but with much finer control on the actions allowed for the program.) On one system, the script could be run with the SETPRV privilege, which is equivalent to setuid root on Unix. This tells an intruder that efforts should be concentrated on this particularly vulnerable server. 4.4.3 Format string bug in PerlRTE_example1.pl See http://wasd.vsm.com.au/ht_root/src/perl/readmore.html for a description of this script and its sources. It contains in particular: printf ("$name=\"$ENV{$name}\"\n"); The variable $name comes from the user and is not filtered, so a classic format string attack is possible. For example: http://wasd.vsm.com.au/plrte/PerlRTE_example1/%25x%25x%25x (where %25 is the hex encoding of the character '%') gives: PATH_INFO="/000" PATH_TRANSLATED="HT_ROOT:[000000]000" REQUEST_URI="/plrte/PerlRTE_example1/ 0 0 0" where the number of zeroes is the number of %x format indicators. This bug is probably not exploitable, but this should be checked. This script also gives away a lot of potentially useful information for an intruder (all logical names). 4.4.4 Potential denial of service in print.com The cgi script print.com allows printing a file on the server from a remote location. This script is enabled by default. The source is available at http://wasd.vsm.com.au/script/print.com This script attempts to restrict the IP addresses allowed to print: $ HPRINTS_ALLOWED = "131.185.250.*" Anyone in this IP range can force the printer to run out of paper. Anyway, it should be fairly easy to spoof the source IP address. 5. Solutions The WASD documentation has a section "Securing the site": http://wasd.vsm.com.au/ht_root/doc/htd/htd_0600.html#98 This is essential reading for any administrator of a WASD server. However some of the recommendations are not effective. In particular the example given in section "Package tree" to block all access to the /ht_root/ tree except for selected areas, does not work correctly for WASD versions up to 8.0. Here is a list of minimum recommendations, which can be put in place even with the existing versions WASD (up to 8.0). Such recommendations might seem obvious for Apache users, but unfortunately many of the WASD sites which were tested didn't follow these basic recommendations. 5.1 Check for new WASD versions at http://wasd.vsm.com.au/WASD/ At the time of writing this advisory, version 8.1 of WASD is in preparation. It will fix all the known security problems, provide from installation a directory structure and associated permissions facilitating minimum necessary upward security adjustments, and use much more restrictive default access rules than previous versions. An advisory written by the WASD author should be available at http://wasd.vsm.com.au/ht_root/doc/misc/wasd_advisory_020925.txt Existing versions 8.0 and 7.2 will each have an update kit available (8.0.1 and 7.2.4 or later). These will include a server with fixes for all known security issues and a script install_secure.com which adjusts the existing directory structure and permissions to conform with that implemented for 8.1. If these updates and/or version 8.1 or later is not available or not applicable to a given site, then check for a kit that provides just the install_secure.com script. This can be used standalone on 7.x and 8.0 installations for significant improvements to site security. If none of these are available or applicable to a given site, then take at least the minimum precautions described below in sections 5.2 to 5.6. 5.2 Use a separate document root, not the whole web server root Put the document root in /ht_root/wwwroot, not /ht_root, with mapping rules such as: pass /* /ht_root/wwwroot/* fail /ht_root/* You can add a rule such as fail /-/* but unfortunately it will not be very effective because it can be trivially bypassed with all versions of the WASD server up to 8.0. 5.3 Use a subdirectory for a user document root, not the home directory Use rules such as: user /~*/* /user_disk/*/web/* redirect /~* /~*/ 5.4 Set file protections and ACLs correctly Since the WASD access restrictions can be bypassed, the only effective protection is that provided by the system itself, OpenVMS. The web server runs by default as user http$server, so make sure that directories supposed to be protected are not readable by this user. Take at least the following minimum precautions: . Make sure that all directories under ht_root are owned by user SYSTEM, not by user HTTP$SERVER. Add specific ACLs for directories and files which must be readable or writable by HTTP$SERVER, but only these: $ set file /owner=system /prot=(s:rwed,o:rwed,g,w) - ht_root:[000000]local.dir, ht_root:[local...]*.*;* , - ht_root:[000000]http$server.dir, ht_root:[http$server...]*.*;* $ set security /acl=((ident=http$server,access=e) /delete=all - ht_root:[000000]local.dir $ set security /acl=((ident=http$server,access=r+e) /delete=all - ht_root:[local]*.com;* $ set security /acl=((ident=http$server,access=r+e) /delete=all - ht_root:[http$server]*.com;* Be particularly careful about ACLs on files in directory [local]. Only the .com files there should be readable by http$server; the rest contains very sensitive information. . To prevent all the different ways of reading scripts, set protection Execute only instead of Read+Execute on all essential scripts (and delete all other scripts as described below). Check at least the directories /script, /script_local, /vax and /axp, plus any other directories mentioned in the logical name CGI-BIN. $ set file /owner=system /prot=(s:rwed,o:rwed,g,w) - ht_root:[000000]script*.dir, ht_root:[script*...]*.*;* , - ht_root:[000000]axp.dir, ht_root:[axp...]*.*;* , - ht_root:[000000]vax.dir, ht_root:[vax...]*.*;* $ set security /acl=((ident=http$server,access=e) /delete=all - ht_root:[000000]script*.dir $ set security /acl=((ident=http$server,access=e) /delete=all - ht_root:[script*]*.*;* $ set security /acl=((ident=http$server,access=e) /delete=all - ht_root:[000000]axp.dir ! or vax.dir depending on architecture $ set security /acl=((ident=http$server,access=e) /delete=all - ht_root:[axp]*.*;* ! or [vax]*.*;* depending on architecture . To prevent reading the web server logs, set ACLs to allow write-only access for user http$server on directory [log] and its subdirectories. . To prevent all the forms of directory traversal, set protection Execute only instead of Read+Execute for directories which must be protected from traversal. You can also add rules in httpd$map.conf fail /tree/ fail /tree/* fail /upd/* fail /where/* fail /query/* fail /extract/* but do not rely only on these rules, set ACLs correctly first. 5.5 Remove all unused cgi programs This is a basic principle, but unfortunately it was not followed in any of the tested sites: remove *all* unused cgi scripts and executables. Some of the programs provided by default with the WASD server are very dangerous. So check at least the directories script, script_local, axp and vax (all under ht_root) and move anything which is unused somewhere outside ht_root, with very strict ACLs. Let me insist again: at least one of the programs provided by default with the WASD server is extremely dangerous. I will not name it to avoid helping too much the script kiddies, so in doubt remove everything unless you know that you absolutely need it. If the machine hosting the web server has some untrusted users with local accounts on the machine, then forbid execution of arbitrary cgi scripts in ~untrusted/cgi-bin with a rule such as: fail /~untrusted/cgi-bin/* put _before_ the rule exec /~*/cgi-bin/* /user_disk/*/cgi-bin/* or better, block cgi access to all local users. 5.6 Never run cgi scripts under the account of a privileged user If you really need privileges in your cgi script, then force the https protocol and force user authentication. Read the WASD documents http://wasd.vsm.com.au/ht_root/doc/htd/htd_1200.html Never do user authentication without SSL (the https protocol). Do not run cgi scripts as user http$server. Strictly speaking, this account is not privileged, but since it is also used to run the main server, faulty cgi scripts can cause more harm than totally unprivileged users. If the machine hosting the web servers has untrusted local users with a ~username/cgi-bin directory, then it is preferable to either remove the ability to run scripts there, or at least run the scripts as the user. Running those scripts as user http$server lets the untrusted users do many things to the web server, such as reading the server logs if ACLs are not set correctly, or killing the main server process; it would be the equivalent of giving the additional right [http$server] to the untrusted users. 6. Examples of site weaknesses Several WASD sites were used for testing the various weaknesses. An actual intrusion was made on two systems only to check that it was indeed feasible. The system administrators were warned immediately. On other systems the presence of the vulnerabilities was checked but was not used for intrusion. 6.1 Site A The first system compromise was not entirely due to weaknesses in WASD and was only possible because of a flaw in an independent cgi script which is not part of the WASD distribution. However without the WASD vulnerabilities, the faulty script would not have been found and its source analysed to find the weakness. The system was rather well (but not perfectly) configured. But the system administrator had left a faulty script in his cgi-bin directory. The home directory was correctly protected by ACLs so even the WASD vulnerabilities didn't allow reading it. The list of scripts and their sources were supposed to be protected by the WASD configuration, but it was very easy to work around this using the "upd" and the "extract" builtin scripts as described in section 4.2.7. There was an another weakness in the WASD configuration: the script was run under the identity of the owner of the script, which was the system administrator. This account was not the predefined SYSTEM account but still had many privileges. It was then possible to remotely run any command with those privileges. The system administrator was told how the compromise was done and how to fix the server. 6.2 Site B The story is very similar, but in this case a flaw in one of the scripts provided by WASD was used. The flaw was particularly severe because the particular configuration used at this site allowed some scripts to be run under the identity of a system administrator, who had all privileges (SETPRV, which is the equivalent of root). The system administrators were warned about the flaw and fixed it immediately. 6.3 Sites C,D,E On three different sites, all directories and files were owned by the user which runs all cgi scripts. So these sites were particularly vulnerable: any flaw in a scrit can immediately lead to a SYSTEM compromise because the script has write access to everything. The system administrators have been warned and the sites have been fixed. 6.4 Almost all sites In almost all tested sites, it was very easy to get read access to important configuration files that are supposed to stay confidential, and to get the list and sources of all cgi scripts, including site specific scripts. 7. Conclusion Do not blindly believe that, if something runs on OpenVMS, then it must be secure. Common sense precautions such as not giving away the entire server tree or all script sources, and checking user input in all cgi scripts, must be taken even on OpenVMS. If you are using WASD 8.0 or earlier, fix your configuration *immediately*. Jean-loup Gailly http://gailly.net PGP or GPG key: http://jl.gailly.net/jloup.asc PGP sig: E3EC F4DF 7EDB E724 A3EC FBC2 D9A2 7D25 0196 71A7 Questions about the WASD server should be sent to its author Mark Daniel . Questions about this advisory should be sent to Jean-loup Gailly 8. Acknowledgments Thanks to Mark Daniel, author of the WASD server, for making it free software, for quickly answering to my initial report, and for working insane hours to fix all the problems I reported. Thanks to Beave and Doc Cypher for providing free accounts on their respective OpenVMS servers, for testing the proposed solutions and suggesting improvements. Thanks to Jeremy Begg for hosting the WASD demonstration server. 9. Document history - Sept. 16, 2002: draft 0.1, WASD vendor contacted - Sept. 16, 2002: draft 0.2, added section "don't run cgi as trusted user" - Sept. 19, 2002: draft 0.3 . mention the possibility of remote SYSTEM compromise . added disclosure of full physical path in 4.2.2 . added directory traversal with glist.com in 4.2.6 . added section 4.3 "CGI scripts are run under the identity of the server" - Sept. 23, 2002: CERT contacted - Sept. 25, 2002: version 1.0. Added reference to WASD advisory in 5.1 - Sept. 25, 2002: Info-WASD mailing list informed - Sept. 26, 2002: sent to Bugtraq for publication This document is available at http://jl.gailly.net/security/wasd-vuln-2002-09.txt It is provided under the policy documented at http://www.wiretrip.net/rfp/policy.html -----BEGIN PGP SIGNATURE----- Version: GnuPG v1.0.6 (GNU/Linux) Comment: For info see http://www.gnupg.org iD8DBQE9kYF02aJ9JQGWcacRAiZSAJ9StmBUBUYnWC+Smnq5ZzC7EsrolQCfYeSX LVJofzQzIUzriquN6J847GM= =F2B9 -----END PGP SIGNATURE-----