Haxx Libcurl

This flaw allows an attacker to insert cookies at will into a running program using libcurl, if the specific series of conditions are met.

libcurl performs transfers. In its API, an application creates "easy handles" that are the individual handles for single transfers.

libcurl provides a function call that duplicates en easy handle called [curl_easy_duphandle](https://curl.se/libcurl/c/curl_easy_duphandle.html).

If a transfer has cookies enabled when the handle is duplicated, the cookie-enable state is also cloned - but without cloning the actual cookies. If the source handle did not read any cookies from a specific file on disk, the cloned version of the handle would instead store the file name as `none` (using the four ASCII letters, no quotes).

Subsequent use of the cloned handle that does not explicitly set a source to load cookies from would then inadvertently load cookies from a file named `none` - if such a file exists and is readable in the current directory of the program using libcurl. And if using the correct file format of course.

This flaw makes curl overflow a heap based buffer in the SOCKS5 proxy handshake.

When curl is asked to pass along the host name to the SOCKS5 proxy to allow that to resolve the address instead of it getting done by curl itself, the maximum length that host name can be is 255 bytes.

If the host name is detected to be longer, curl switches to local name resolving and instead passes on the resolved address only. Due to this bug, the local variable that means "let the host resolve the name" could get the wrong value during a slow SOCKS5 handshake, and contrary to the intention, copy the too long host name to the target buffer instead of copying just the resolved address there.

The target buffer being a heap based buffer, and the host name coming from the URL that curl has been told to operate with.

An authentication bypass vulnerability exists in libcurl prior to v8.0.0 where it reuses a previously established SSH connection despite the fact that an SSH option was modified, which should have prevented reuse. libcurl maintains a pool of previously used connections to reuse them for subsequent transfers if the configurations match. However, two SSH settings were omitted from the configuration check, allowing them to match easily, potentially leading to the reuse of an inappropriate connection.

A double free vulnerability exists in libcurl <8.0.0 when sharing HSTS data between separate "handles". This sharing was introduced without considerations for do this sharing across separate threads but there was no indication of this fact in the documentation. Due to missing mutexes or thread locks, two threads sharing the same HSTS data could end up doing a double-free or use-after-free.

An authentication bypass vulnerability exists libcurl <8.0.0 in the connection reuse feature which can reuse previously established connections with incorrect user permissions due to a failure to check for changes in the CURLOPT_GSSAPI_DELEGATION option. This vulnerability affects krb5/kerberos/negotiate/GSSAPI transfers and could potentially result in unauthorized access to sensitive information. The safest option is to not reuse connections if the CURLOPT_GSSAPI_DELEGATION option has been changed.

An authentication bypass vulnerability exists in libcurl <8.0.0 in the FTP connection reuse feature that can result in wrong credentials being used during subsequent transfers. Previously created connections are kept in a connection pool for reuse if they match the current setup. However, certain FTP settings such as CURLOPT_FTP_ACCOUNT, CURLOPT_FTP_ALTERNATIVE_TO_USER, CURLOPT_FTP_SSL_CCC, and CURLOPT_USE_SSL were not included in the configuration match checks, causing them to match too easily. This could lead to libcurl using the wrong credentials when performing a transfer, potentially allowing unauthorized access to sensitive information.

When sending data to an MQTT server, libcurl <= 7.73.0 and 7.78.0 could in some circumstances erroneously keep a pointer to an already freed memory area and both use that again in a subsequent call to send data and also free it *again*.

libcurl keeps previously used connections in a connection pool for subsequenttransfers to reuse, if one of them matches the setup.Due to errors in the logic, the config matching function did not take 'issuercert' into account and it compared the involved paths *case insensitively*,which could lead to libcurl reusing wrong connections.File paths are, or can be, case sensitive on many systems but not all, and caneven vary depending on used file systems.The comparison also didn't include the 'issuer cert' which a transfer can setto qualify how to verify the server certificate.

curl 7.63.0 to and including 7.75.0 includes vulnerability that allows a malicious HTTPS proxy to MITM a connection due to bad handling of TLS 1.3 session tickets. When using a HTTPS proxy and TLS 1.3, libcurl can confuse session tickets arriving from the HTTPS proxy but work as if they arrived from the remote server and then wrongly "short-cut" the host handshake. When confusing the tickets, a HTTPS proxy can trick libcurl to use the wrong session ticket resume for the host and thereby circumvent the server TLS certificate check and make a MITM attack to be possible to perform unnoticed. Note that such a malicious HTTPS proxy needs to provide a certificate that curl will accept for the MITMed server for an attack to work - unless curl has been told to ignore the server certificate check.

curl 7.1.1 to and including 7.75.0 is vulnerable to an "Exposure of Private Personal Information to an Unauthorized Actor" by leaking credentials in the HTTP Referer: header. libcurl does not strip off user credentials from the URL when automatically populating the Referer: HTTP request header field in outgoing HTTP requests, and therefore risks leaking sensitive data to the server that is the target of the second HTTP request.

curl 7.41.0 through 7.73.0 is vulnerable to an improper check for certificate revocation due to insufficient verification of the OCSP response.

curl 7.21.0 to and including 7.73.0 is vulnerable to uncontrolled recursion due to a stack overflow issue in FTP wildcard match parsing.

Due to use of a dangling pointer, libcurl 7.29.0 through 7.71.1 can use the wrong connection when sending data.

A heap buffer overflow in the TFTP receiving code allows for DoS or arbitrary code execution in libcurl versions 7.19.4 through 7.64.1.

libcurl versions from 7.34.0 to before 7.64.0 are vulnerable to a heap out-of-bounds read in the code handling the end-of-response for SMTP. If the buffer passed to `smtp_endofresp()` isn't NUL terminated and contains no character ending the parsed number, and `len` is set to 5, then the `strtol()` call reads beyond the allocated buffer. The read contents will not be returned to the caller.

libcurl versions from 7.36.0 to before 7.64.0 are vulnerable to a stack-based buffer overflow. The function creating an outgoing NTLM type-3 header (`lib/vauth/ntlm.c:Curl_auth_create_ntlm_type3_message()`), generates the request HTTP header contents based on previously received data. The check that exists to prevent the local buffer from getting overflowed is implemented wrongly (using unsigned math) and as such it does not prevent the overflow from happening. This output data can grow larger than the local buffer if very large 'nt response' data is extracted from a previous NTLMv2 header provided by the malicious or broken HTTP server. Such a 'large value' needs to be around 1000 bytes or more. The actual payload data copied to the target buffer comes from the NTLMv2 type-2 response header.

libcurl versions from 7.36.0 to before 7.64.0 is vulnerable to a heap buffer out-of-bounds read. The function handling incoming NTLM type-2 messages (`lib/vauth/ntlm.c:ntlm_decode_type2_target`) does not validate incoming data correctly and is subject to an integer overflow vulnerability. Using that overflow, a malicious or broken NTLM server could trick libcurl to accept a bad length + offset combination that would lead to a buffer read out-of-bounds.

curl before version 7.61.1 is vulnerable to a buffer overrun in the NTLM authentication code. The internal function Curl_ntlm_core_mk_nt_hash multiplies the length of the password by two (SUM) to figure out how large temporary storage area to allocate from the heap. The length value is then subsequently used to iterate over the password and generate output into the allocated storage buffer. On systems with a 32 bit size_t, the math to calculate SUM triggers an integer overflow when the password length exceeds 2GB (2^31 bytes). This integer overflow usually causes a very small buffer to actually get allocated instead of the intended very huge one, making the use of that buffer end up in a heap buffer overflow. (This bug is almost identical to CVE-2017-8816.)

The URL percent-encoding decode function in libcurl before 7.51.0 is called `curl_easy_unescape`. Internally, even if this function would be made to allocate a unscape destination buffer larger than 2GB, it would return that new length in a signed 32 bit integer variable, thus the length would get either just truncated or both truncated and turned negative. That could then lead to libcurl writing outside of its heap based buffer.

In curl and libcurl 7.52.0 to and including 7.53.1, libcurl would attempt to resume a TLS session even if the client certificate had changed. That is unacceptable since a server by specification is allowed to skip the client certificate check on resume, and may instead use the old identity which was established by the previous certificate (or no certificate). libcurl supports by default the use of TLS session id/ticket to resume previous TLS sessions to speed up subsequent TLS handshakes. They are used when for any reason an existing TLS connection couldn't be kept alive to make the next handshake faster. This flaw is a regression and identical to CVE-2016-5419 reported on August 3rd 2016, but affecting a different version range.

libcurl 7.49.0 to and including 7.57.0 contains an out bounds read in code handling HTTP/2 trailers. It was reported (https://github.com/curl/curl/pull/2231) that reading an HTTP/2 trailer could mess up future trailers since the stored size was one byte less than required. The problem is that the code that creates HTTP/1-like headers from the HTTP/2 trailer data once appended a string like `:` to the target buffer, while this was recently changed to `: ` (a space was added after the colon) but the following math wasn't updated correspondingly. When accessed, the data is read out of bounds and causes either a crash or that the (too large) data gets passed to client write. This could lead to a denial-of-service situation or an information disclosure if someone has a service that echoes back or uses the trailers for something.

curl and libcurl before 7.57.0 on 32-bit platforms allow attackers to cause a denial of service (out-of-bounds access and application crash) or possibly have unspecified other impact because too little memory is allocated for interfacing to an SSL library.

The FTP wildcard function in curl and libcurl before 7.57.0 allows remote attackers to cause a denial of service (out-of-bounds read and application crash) or possibly have unspecified other impact via a string that ends with an '[' character.

The NTLM authentication feature in curl and libcurl before 7.57.0 on 32-bit platforms allows attackers to cause a denial of service (integer overflow and resultant buffer overflow, and application crash) or possibly have unspecified other impact via vectors involving long user and password fields.

An IMAP FETCH response line indicates the size of the returned data, in number of bytes. When that response says the data is zero bytes, libcurl would pass on that (non-existing) data with a pointer and the size (zero) to the deliver-data function. libcurl's deliver-data function treats zero as a magic number and invokes strlen() on the data to figure out the length. The strlen() is called on a heap based buffer that might not be zero terminated so libcurl might read beyond the end of it into whatever memory lies after (or just crash) and then deliver that to the application as if it was actually downloaded.