You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@mina.apache.org by "Thomas Wolf (JIRA)" <ji...@apache.org> on 2018/11/18 12:56:00 UTC
[jira] [Comment Edited] (SSHD-868) Add some protection against
maliciously crafted packets
[ https://issues.apache.org/jira/browse/SSHD-868?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16690892#comment-16690892 ]
Thomas Wolf edited comment on SSHD-868 at 11/18/18 12:55 PM:
-------------------------------------------------------------
Is the length encoding really an issue? Ssh packet length is limited, in Apache sshd apparently to 256kB (see [{{AbstractSession.decode()}}, line 1007|https://github.com/apache/mina-sshd/blob/0241783/sshd-core/src/main/java/org/apache/sshd/common/session/helpers/AbstractSession.java#L1007]). And for instance {{ByteArrayBuffer.getString(Charset)}} checks that the buffer does have enough unread bytes available. So if someone sends a packet containing a bogus string with a length 0xffffffff, sshd will throw a buffer exception, but it won't run into an OOM.
The number of challenges in keyboard-interactive authentication might perhaps be a problem. A malicious server could just keep on sending zero-prompt info requests to make the client be stuck. At least until the client-side authentication timeout kicks in? Perhaps counting the zero-prompt info requests separately and aborting when that count reachesĀ {{2 * maxTrials}} might be an idea. I don't see any such count in openssh, though. They seem to count only the number of inits for keyboard-interactive. See [{{sshconnect2.c}}, lines 1705ff|https://github.com/openssh/openssh-portable/blob/aaed635/sshconnect2.c#L1705]. Looks to me that the Apache sshd way of counting is better.
was (Author: wolft):
Is the length encoding really an issue? Ssh packet length is limited, in Apache sshd apparently to 256kB (see [{{AbstractSession.decode()}}, line 1007|https://github.com/apache/mina-sshd/blob/0241783/sshd-core/src/main/java/org/apache/sshd/common/session/helpers/AbstractSession.java#L1007]). And for instance {{ByteArrayBuffer.getString(Charset)}} checks that the buffer does have enough unread bytes available. So if someone sends a packet containing a bogus string with a length 0xffffffff, sshd will throw a buffer exception, but it won't run into an OOM.
The number of challenges in keyboard-interactive authentication might perhaps be a problem. A malicious server could just keep on sending zero-prompt info requests to make the client be stuck. Perhaps counting the zero-prompt info requests separately and aborting when that count reachesĀ {{2 * maxTrials}} might be an idea. I don't see any such count in openssh, though. They seem to count only the number of inits for keyboard-interactive. See [{{sshconnect2.c}}, lines 1705ff|https://github.com/openssh/openssh-portable/blob/aaed635/sshconnect2.c#L1705]. Looks to me that the Apache sshd way of counting is better.
> Add some protection against maliciously crafted packets
> -------------------------------------------------------
>
> Key: SSHD-868
> URL: https://issues.apache.org/jira/browse/SSHD-868
> Project: MINA SSHD
> Issue Type: Improvement
> Affects Versions: 2.1.1
> Reporter: Goldstein Lyor
> Assignee: Goldstein Lyor
> Priority: Major
> Labels: memory, overflow, security
> Fix For: 2.1.1
>
>
> According to [RFC4256 - section 3.2|https://tools.ietf.org/html/rfc4256]
> {quote}
> The server SHOULD take into consideration that some clients may not
> be able to properly display a long name or prompt field (see next
> section), and limit the lengths of those fields if possible.
> {quote}
> The current code in {{UserAuthKeyboardInteractive#processAuthDataRequest}} does not make sure that the number of challenges or the length of each challenge is reasonable (not to mention the other packet components). Therefore, a maliciously crafted packet can cause out-of-memory errors by requesting an extremely large number of responses or sending very large challenges.
> It is important to notice that this problem is not limited to the {{keyboard-interactive}} protocol but to the entire packet encode/decode mechanism since it is a RLE (read-length encoding). Wherever possible we should add some reasonable but large enough limitations on the expected size of strings/arrays/etc.. being decoded from incoming SSH packets.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)