CVE-2020-8444: analysisd: OS_ReadMSG heap use-after-free with ossec-alert msgs.

[cpu]

The ossec-analysisd's OS_ReadMSG function calls OS_CleanMSG at the start of processing a received message from the ossec queue UNIX domain socket.

In src/analysisd/cleanevent.c the OS_CleanMSG function populates the lf struct, setting fields like log, hostname and program_name to substrings of the lf->full_log buffer.

After cleaning any messages that meet the ossec-alert decoder's criteria are given to that decoder for further processing.

After processing an ossec alert msg from a client the ossec alert decoder will free the lf->full_log pointer at the end of its processing, replacing it with a new pointer and populating lf->generated_rule:

ossec-hids/src/analysisd/decoders/plugins/ossecalert_decoder.c

Lines 184 to 191 in abb36d4

	free(lf->full_log);
	lf->full_log = NULL;
	os_strdup(tmpstr_buffer, lf->full_log);
	lf->log = lf->full_log;
	/* Rule that generated. */
	lf->generated_rule = rule_pointer;

Though the OSSECAlert_Decoder_Exec function returns NULL and not 1 further rule processing of the lf struct occurs during OS_ReadMSG because of the lf->generated_rule set by the decoder before freeing lf->full_log.

ossec-hids/src/analysisd/analysisd.c

Lines 872 to 879 in abb36d4

	if (lf->decoder_info->type == OSSEC_ALERT) {
	if (!lf->generated_rule) {
	goto CLMEM;
	}
	/* Process the alert */
	currently_rule = lf->generated_rule;
	}

If any subsequent processing associated with the generated rule accesses the lf->hostname or lf->program_name fields set by OS_CleanMSG they will be accessing memory of a freed heap chunk previously containing the lf->full_log.

I believe the bug was introduced in fcca013 on July 23, 2008 and affects OSSEC v2.7+.

This is triggerable via an authenticated client through the ossec-remoted. The client needs only write a ossecalert message that will have the program_name or hostname set during OS_CleanMSG.

I don't have a strong sense for the possibility of exploitation. I suspect this may be turned into an out of bounds read of heap memory accessing program_name or hostname during rule processing if the area pointed to after the syscheck decoder free isn't null terminated.

One possible fix would be for the ossecalert decoder to os_strdup the lf->hostname and lf->program_name before freeing full_log.

[cpu]

One possible fix would be for the ossecalert decoder to os_strdup the lf->hostname and lf->program_name before freeing full_log.

thinking about this more: I think this proposed fix would introduce a memory leak. Using os_strdup will mean that the hostname and program_name pointers no longer point into the lf->log or lf->full_log buffer and will instead point to newly allocated memory.

The Free_Eventinfo function seems to be written with the assumption that freeing lf->full_log will free the program_name and hostname. Unlike other fields of the Eventinfo struct these two fields are not explicitly freed in the body of Free_Eventinfo.

This is likely a case where someone more familiar with this codebase will have to suggest a better fix.

[cpu]

This was assigned CVE-2020-8444

[NicoleG25]

Is there any progress on this or some sort of ETA ?
Thanks in advance :)
@atomicturtle

[attritionorg]

Bump as well. I don't see a fixing commit for this one but not sure if I overlooked something.