In our 2021 blog post, we focused on identifying quick wins for optimizing Windows Events, and provided a free spreadsheet (really free, not even a regwall) that indicated Windows Events that could be safely ignored, some of which cost lots for SIEM engines to ingest. This post takes a broader Windows Audit Policy view, and offers another free resource – this time taking a broader look in the context of comparing your setup for Windows Audit Policy, and the venerable CIS Benchmark for Windows 2019 Server.
If there’s sufficient interest i’ll follow up with a development effort for a Python tool (also freely available, on Github) that connects to your Windows server and performs the CIS Benchmark assessment as indicated in the spreadsheet.
SIEM Nightmares
Based on many first hand observations and second hand accounts, it’s not a stretch to say that many organisations are suffering from SIEM configuration issues, for which the result is a low signal-to-noise ratio. Your SIEM is ingesting lots of events, many of which are not at all helpful, and with most vendors charging by volume, it gets expensive. At the same time, the false negative problem is all too common. Forensics investigations reveal all too often that there are no events recorded by the expensive SIEM, that even closely relate to the incident. I hope you are never in this scenario. The short-term impact is never good.
Taking SIEM as a capability, if one is to advise on how to improve things, it is rarely ever about the technology. When one asks Analysts (and based on job postings, also hiring managers) about SIEM, it’s clear the first thing that comes to mind is Splunk. ELK, Sentinel, etc. I would estimate the technology-only focus with SIEM to be the norm rather than the exception, and it comes hand-in-hand with a failure to detect privilege elevations, and lateral movements for example.
There are some advisories that we can give out that are independent of your architecture, but many questions about SIEM configuration can only be answered by you, using your knowledge of the IT landscape in your organisation. The advisories in the referenced spreadsheet cover the “noise” part of the signal-to-noise ratio. These are events that are sure to be noise to at least a 90% level of assurance, from a security perspective.
Addtional Context on the Spreadsheet
Some context around the spreadsheet: where there is a CIS Benchmark metric for a specific Audit Subcategory, the spreadsheet follows exactly the CIS recommended setting. But there are some (e.g. DS Access –> Directory Service Access) where this subcategory was not covered by CIS. In these cases, an assessment is made based on our real-experience observations of logging volumes, versus the security (not the IT diagnostic, or other value) value of Audit Subcategories. In this case of the Directory Service Access subcategory, it can be turned off from a security perspective.
There is limited information available regarding actual experiences with specific event ID volumes. In 2018, I had the opportunity to track Windows events in a Splunk architecture for a government department. During this time, I recorded the occurrences of events over a 24-hour period on a network of approximately 150 Windows servers of various versions, some of which were quite exotic. This information has been valuable in supporting decisions related to whether or not to disable auditing.
SIEM Forwarder Filtering
There is another option offered by some SIEM vendors and that is to filter events by Event ID. Overall, the more resource-friendly approach is to prevent the events being generated at source, but in many cases this may not be feasible. Splunk for example allows you to filter at forwarders (via the inputs.conf file on the Splunk forwarder. This file is usually located in the $SPLUNK_HOME/etc/system/local/ directory … more info – BTW it looks like Splunk agrees with us on the 4662 event mentioned as an example above. Yay!).
Credits and Disclaimers
Windows Events are sometimes tricky to understand, both with respect of what the developers intended with those events, and the conditions under which they are generated. Sometimes with Windows Events, we are completely in unknown territory, even if there is some Microsoft documentation that covers them. Here’s one example from Microsoft documentation to fill us with confidence – “This auditing subcategory should not have any events in it, but for some reason Success auditing will enable the generation of event 4985″.
Ultimately only you can decide what’s best for the health of your SOC/SIEM. Only you know your network and your applications. The document supplied here was only intended as a guide, and to aid decision making. It was not intended to make decisions for you.
The cybersecurity landscape often focuses on the more sensational aspects, such as high-profile hacks or fake influencers, which can overshadow the essential work done by countless professionals in the background. These unsung heroes are dedicated to ensuring the stability and security of our digital infrastructure, and their contributions should not be underestimated. Among those are tthe likes of Randy Franklin Smith (founder of Ultmate Windows Security) who has put together an “encyclopedia” of Windows Event IDs. The experiences shared there were used in-part to form a view on whether or not to reject or accept certain Windows Events.
It’s clear from social media and first hand reports, that the awareness of what VA (Vulnerability Assessment) scanners are really doing in testing scenarios is quite low. So I setup up a test box with Ubuntu 18 and exposed some services which are well known to the hacker community and also still popular in production business use cases: Secure Shell (SSH) and an Apache web service.
This post isn’t an attack on VA products at all. It’s aimed at setting a more healthy expectation, and I will cover a test scenario with a packet sniffer (Wireshark), Nessus Professional, and OpenVAS, that illustrates the point.
I became aware 20 years ago, from validating VA scanner output, that a lot of what VA scanners barf out is alarmist (red flags, CRITICAL [fix NOW!]) and also based purely on guesswork – when the scanner “sees” a service, it grabs a service banner (e.g. “OpenSSH 7.6p1 Ubuntu 4ubuntu0.3”), looks in its database for public disclosed vulnerability with that version, and flags vulnerability if there are any associated CVEs. Contrary to popular belief, there is no actual interaction in the way of further investigating or validating vulnerability. All vulnerability reporting is based on the service banner. So if i change my banner to “hi OpenVAS”, nothing will be reported. And in security, we like to advise hiding product names and versions – this helps with drive-by style automated attacks, in a much more effective way than for example, changing default service ports.
This article then demonstrates the VA scanner behaviour described above and covers developments over the past 20 years (did things improve?) with the two most commonly found scanners: Nessus and OpenVAS, which even if are not used directly, are used indirectly (vendors in this space do not recreate the wheel, they take existing IP – all legal I’m sure – and create their own UI for it). It was fairly well-known that Nessus was the basis of most commercial VAs in the 00s, and it seems unlikely that scenario has changed a great deal.
Test Setup
So if I look at my test box setup I see from port scan results (nmap):
PORT STATE SERVICE VERSION 22/tcp open ssh OpenSSH 7.6p1 Ubuntu 4ubuntu0.3 (Ubuntu Linux; protocol 2.0) 25/tcp open smtp Postfix smtpd 80/tcp open http Apache httpd 2.4.29 ((Ubuntu)) 139/tcp open netbios-ssn Samba smbd 3.X - 4.X (workgroup: WORKGROUP) 445/tcp open netbios-ssn Samba smbd 3.X - 4.X (workgroup: WORKGROUP) 3000/tcp open http Apache httpd 2.4.29 ((Ubuntu)) 5000/tcp open http Docker Registry (API: 2.0) 8000/tcp open http Apache httpd 2.4.29
So…naughty, naughty. Apache is not so old but still I’d expect to see some CVEs flagged, and I can say the same for the SSH service. Samba is there too in a default format. Samba is Linux’s implementation of MS Windows SMB (Server Message Block) and is full of holes. The Postfix mail service is also quite old, and there’s a Docker API exposed! All this would get an attacker quite excited, and indeed there’s plenty of automated attack scenarios which would work here.
There was also an EOLPhpmyadmin and EOL jQuery wrapped up in the web service.
Developments in Two Decades
So there has been some changes. For want of a better word, there’s now more honesty. In the case of OpenVAS, for vulnerability that involves grabbing a banner and assuming vulnerability based on this, there is a Quality of Detection (QoD) rating, which is set as default at around 70%. This is a kind of probability rating for a finding not being a false positive. Interestingly those findings that involve a banner grab are way down there under 50, and most are no longer flagged as “critical”.
Nessus, for its banner-grabbed vulnerabilities, is more explicit and it is report will state “Note that Nessus has not tested for this issue but has instead relied only on the application’s self-reported version number.”
Even 7 years ago, there would be lots of issues reported for an outdated Apache or SSH service, many of which would be flagged wrongly as CRITICAL, but not necessarily exploitable, and the existance of the vulnerability was based only on a text banner. So these more recent VA versions are an improvement, but its clear the awareness out there of these issues is still quite low. The problem is now – we do want to see if services are downlevel, so please $VENDOR, don’t hide them (more on this later).
First Scan – Banners On Display
So using Wireshark, sniffing HTTP on port 80 (plain text) we have the following…
The packets highlighted in black are the only two of any interest, wherein OpenVAS has used the HTTP GET method to request for “/”, and receives a response where the header shows the product (Apache) and version (2.4.29).
Note the Wireshark filter used (tcp.port == 80 and http). Other than the initial exchange where a banner was grabbed, there was no further interaction. This was the same for Nessus.
What was reported? Well, for OpenVAS, a handful of potential CVEs were reported but I had to lower the QoD to see them! Which is interesting. If anything this is moving the bar too far in the opposite direction. I mean as an owner of this system, I do want to know if i am running old warez!
For Nessus, 6 Apache CVEs were reported with either critical or “high” severity. Overall, I had a similar experience with that of OpenVAS except to even see the Apache issues reported I had to beg the scanner with the following scan configuration setup:
Settings –> Assessment –> Override normal accuracy and show potential false alarms
Settings –> Assessment –> perform thorough tests
Settings –> Advanced –> enable safe checks on (and i also tried the “off” option)
Settings –> Advanced –> plugins –> web servers –> enabled. This is the Apache vulnerability section
For the SSH service, OpenVAS reported 3 medium issues which is roughly what i was expecting. Nessus did not report any at all! Answers on a postcard for that one.
Banners Concealed
What was interesting was that the Secure Shell service doesn’t present an option to hide the banner any more, and on investigation, the majority-held community-version of this story is that the banner is needed in some cases.
Apache however did present a banner obfuscation option. For Ubuntu 18 and Apache 2.4.29, this involved:
apt install libapache2-mod-security2
a2enmod security2
edit /etc/apache2/conf-available/security.conf
ServerTokens set to “Prod”
systemctl restart apache2
This setup results in the following banner for Apache: Apache httpd – so no version number.
The outcome? As expected, all mention of Apache has now ended. Neither OpenVAS or Nessus reported anything to do with Apache of any note.
What DID The Scanners Find?
Just to summarise the findings when the banners were fully on display…it wasn’t a blank slate. There were some findings. Here are the highlights – for OpenVAS:
All Critical issues detected were related to PHPMyAdmin, plus one related to jQuery being EOL, but not stating any particular vulnerability. These version numbers are remotely queriable and this is the basis on which these issues were reported.
The SSH and Apache issues.
Other lower criticality issues were around certificate ciphers.
Some CVSS 6, medium issues with Samba – again these are banner-grabbed guesswork findings.
Nessus didn’t report anything outside of what OpenVAS flagged. OpenVAS reported significantly more issues.
It should be said that both scanners did a lot of querying for HTTP application layer issues that could be seen in the packet sniffer output. For example, queries were made for Python/Django settings.py (database password), and other HTTP gotchas.
Unauthenticated Versus Credentialed Testing
With VA Scanners, the picture hasn’t really changed in 20 years. If anything the picture is worse now because the balance with banner-grabbing guesswork has swung too far the other way, and we have to plead with the scanners to tell us about downlevel software versions. This is presumably an effort to reduce the number of false positives, but its not an advisable strategy. It’s perfectly ok to let us know we are running old wares and if we want, we should be able to see the CVEs associated with our listening services, even if many of them are false positives (and I can say from 20 years of network penetration testing, there will be plenty).
With this type of unauthenticated VA scanning though, the real problem has always been false negatives (to the extent that an open Docker API wasn’t flagged as a problem by either scanner), but none of the other commercial tools out there (I have tried a few in recent years) will be in a better position, because there is hard-limit that can be achieved non-locally with no adminstrative authentication credentials.
Both Nessus and OpenVAS allow use of credentialled based testing but its clear this aspect was never a part of the core design. Nessus has expanded its portfolio of credentialed tests but in the time allocated I could not get it to work with SSH public key authentication. In any case, a CIS benchmark approach will always be not-so-great, for reasons outside the scope of this article. We also have to be careful about where authentication credentials are stored. In the case of SSH keys, this means storing a private key, and with some vendors the key will be stored in their cloud somewhere out there.
Conclusion
This post focusses on one major aspect of VA scanning that is grabbing banners and reporting on vulnerability based on the findings from the banner. This is better than nothing but its futility is hopefully illustrated here, and this approach is core to most of what VA scanners do for us.
The market priority has always been towards unauthenticated scanning. Little focus was ever given to credentialed scanning. This has to change because the unauthenticated approach is like trying to diagnose a problem with your car without ever lifting the bonnet/hood, and moreover we could be moving into an era where accreditation bodies mandate credentialed scanning.
Part Five – Cryptography and Key Management, and Identity Management
Part Six – Trust (network controls, such as firewalls and proxies), and Resilience
Logging
Notice “Logging” is used here, not “SIEM”. With use of “SIEM”, there is often a mental leap, or stumble, towards a commercial solution. But there doesn’t necessarily need to be a commercial solution. This post invites the reader to take a step back from the precipice of engaging with vendors, and check first if that journey is one you want to make.
Unfortunately, in 2020, it is still the case that many fintechs are doing one of two things:
Procuring a commercial solution without thinking about what is going to be logged, or thinking about the actual business goals that a logging solution is intended to achieve.
Just going with the Cloud Service Provider’s (CSP) SaaS offering – e.g. Stackdriver (now called “Operations”) for Google Cloud, or Security Center for Azure.
Design Process
The process HLD takes into risks from threat modelling (and maybe other sources), and another input from compliance requirements (maybe security standards and legal requirements), and uses the requirements from the HLD to drive the LLD. The LLD will call out the use cases and volume requirements that satisfy the HLD requirements – but importantly, it does not cover the technological solution. That comes later.
The diagram above calls out Splunk but of course it doesn’t have to be Splunk.
Security Operations
The end goal of the design process is heavily weighted towards a security operations or protective monitoring capability. Alerts will be specified which will then be configured into the technological solution (if it supports this). Run-books are developed based on on-going continuous improvement – this “tuning” is based on adjusting to false positives mainly, and adding further alerts, or modifying existing alerts.
The decision making on how to respond to alerts requires intimate knowledge of networks and applications, trust relationships, data flows, and the business criticality of information assets. This is not a role for fresh graduates. Risk assessment drives the response to an alert, and the decision on whether or not to engage an incident response process.
General IT monitoring can form the first level response, and then Security Operations consumes events from this first level that are related to potential security incidents.
Two main points relating this SecOps function:
Outsourcing doesn’t typically work when it comes to the 2nd level. Outsourcing of the first level is more likely to be cost effective. Dr Anton Chuvakin’s post on what can, and cannot be outsourced in security is the most well-rounded and realistic that i’ve seen. Generally anything that requires in-house knowledge and intimacy of how events relate to business risks – this cannot be outsourced at all effectively.
The maturity of SecOps doesn’t happen overnight. Expect it to take more than 12 months for a larger fintech with a complex cloud footprint.
The logging capability is the bedrock of SecOps, and how it relates to other security capabilities can be simplified as in the diagram below. The boxes on the left are self-explanatory with the possible exception of Active Trust Management – this is heavily network-oriented and at the engineering end of the rainbow, its about firewalls, reverse and forward proxies mainly:
Custom Use Cases
For the vast majority of cases, custom use cases will need to be formulated. This involves building a picture of “normal”, so as to enable alerting on abnormal. So taking the networking example: what are my data flows? Take my most critical applications – what are source and destination IP addresses, and what is the port on the server-side of the client-server relationship? So then a possible custom use case could be: raise an alert when a connection is aimed at the server from anywhere other than the client(s).
Generic use cases are no-brainers. Examples are brute force attempts and technology or user behaviour-specific use cases. Some good examples are here. Custom use cases requires an understanding of how applications, networks, and operating systems are knitted together. But both custom and generic use cases require a log source to be called out. For network events, this will be a firewall as the best candidate. It generally makes very little sense to deploy network IDS nodes in cloud.
So for each application, generate a table of custom use cases, and identify a log source for each. Generic use cases are those configured auto-tragically in Splunk Enterprise Security for example. But even Splunk cannot magically give you custom use cases, or even ensure that all devices are included in the coverage for generic use cases. No – humans still have a monopoly over custom use cases and well, really, most of SIEM configuration. AI and Cyberdyne Systems won’t be able to get near custom use cases in our lifetimes, or ever, other than the fantasy world of vendor Powerpoint slides.
Don’t forget to test custom use case alerting. So for network events, spin up a VM in a centrally trusted area, like a management Vnet/VPC for example. Port scan from there to see if alerts are triggered. Netcat can be very useful here too, for spoofing source addresses for example.
Correlation
Correlation was the phrase used by vendors in the heady days of the 00s. The premise was something like this: event A, event B, and event C. Taken in isolation (topical), each seem innocuous. But bake them together and you have a clear indicator that skullduggery is afoot.
I suggest you park correlation in the early stage of a logging capability deployment. Maybe consider it for down the road, once a decent level of maturity has been reached in SecOps, and consider also that any attempt to try and get too clever can result in your SIEM frying circuit boards. The aim initially should be to reduce complexity as much as possible, and nothing is better at adding complexity than correlation. Really – basic alerting on generic and custom use cases gives you most of the coverage you need for now, and in any case, you can’t expect to get anywhere near an ideal state with logging.
SaaS
Operating system logs are important in many cases. When you decide to SaaS a solution, note that you lose control over operating system events. You cannot turn off events that you’re not interested in (e.g. Windows Object auditing events which have had a few too many pizzas). This can be a problem if you decide to go with a COTS where licensing costs are based on volume of events. Also, you cannot turn on OS events that you could be interested in. The way CSPs play here is to assume everything is interesting, which can get expensive. Very expensive.
Note – its also, in most cases, not such a great idea to use a SaaS based SIEM. Why? Because this function has connectivity with everything. It has trust relationships with dev/test, pre-prod, and production. You really want full control over this platform (i.e. be able to login with admin credentials and take control of the OS), especially as it hosts lots of information that would be very interesting for attackers, and is potentially the main target for attackers, because of the trust relationships I mentioned before.
So with SaaS, its probably not the case that you are missing critical events. You just get flooded. The same applies to 3rd party applications, but for custom, in-house developed applications, you still have control of course of the application layer.
Custom, In-house Developed Applications
You have your debugging stream and you have your application stream. You can assign critical levels to events in your code (these are the classic syslog severity levels). The application events stream is critical. From an application security perspective, many events are not immediately intuitively of interest, but by using knowledge of how hackers work in practice, security can offer some surprises here, pleasant or otherwise.
If you’re a developer, you can ease the strain on your infosec colleagues by using consistent JSON logging keys across the board. For example, don’t start with ‘userid’ and then flip to ‘user_id’ later, because it makes the configuration of alerting more of a challenge than it needs to be. To some extent, this is unavoidable, because different vendors use different keys, but every bit helps. Note also that if search patterns for alerting have to cater for multiple different keys in JSON documents, the load on the SIEM will be unnecessarily high.
It goes without saying also: think about where your application and debug logs are being transmitted and stored. These are a source of extremely valuable intelligence for an attacker.
The Technology
The technological side of the logging capability isn’t the biggest side. The technology is there to fulfil a logging requirement, it is not in itself the logging capability. There are also people and processes around logging, but its worth talking about the technology.
What’s more common than many would think – organisation acquires a COTS SIEM tool but the security engineers hate it. Its slow and doesn’t do much of any use. So they find their own way of aggregating network-centralised events with a syslog bucket of some description. Performance is very often the reason why engineers will be grep’ing over syslog text files.
Whereas the aforementioned sounds ineffective, sadly its more effective than botched SIEM deployments with poorly designed tech. It also ticks the “network centralised logging” box for auditors.
The open-source tools solution can work for lots of organisations, but what you don’t get so easily is the real-time alerting. The main cost will be storage. No license fees. Just take a step back, and think what it is you really want to achieve in logging (see the design process above). The features of the open source logging solution can be something like this:
Rsyslog is TCP and covers authentication of hosts. Rsyslog is a popular protocol because it enables TCP layer transmission from most log source types (one exception is some Cisco network devices and firewalls), and also encryption of data in transit, which is strongly recommended in a wide open, “flat” network architecture where eavesdropping is a prevalent risk.
Even Windows can “speak” rsyslog with the aid of a local agent such as nxlog.
There are plenty of Host-based Intrusion Detection System (HIDS) agents for Linux and Windows – OSSEC, Suricata, etc.
Intermediate network logging Rsyslog servers can aggregate logs for network zones/subnets. There are the equivalent of Splunk forwarders or Alienvault Sensors. A cron job runs an rsync over Secure Shell (SSH), which uploads the batches of events data periodically to a Syslog Lake, for want of a better phrase.
The folder structure on the Syslog server can reflect dates – years, months, days – and distinct files are named to indicate the log source or intermediate server.
Splunk sales people have a dart board with my picture on it. To be fair, the official Splunk line is that they want to help their customers save events indexing money because it benefits them in the longer term. And they’re right, this does work for Splunk and their customers. But many of the resellers are either lacking the skills to help, or they are just interested in a quick and dirty install. “Live for today, don’t worry about tomorrow”.
Splunk really is a Lamborghini, and the few times when i’ve been involved in bidding beauty parades for SIEM, Splunk often comes out cheaper believe it or not. Splunk was made for logging and was engineered as such. Some of the other SIEM engines are poorly coded and connect to a MySQL database for example, whereas Splunk has its own database effectively. The difference in performance is extraordinary. A Splunk search involving a complex regex with busy indexers and search heads takes a fraction of the time to complete, compared with a similar scenario from other tools on the same hardware.
Three main ways to reduce events indexing costs with Splunk:
Root out useless events. Windows is the main culprit here, in particular Auditing of Objects. Do you need, for example, all that performance monitoring data? Debug events? Firewall AND NIDS events? Denied AND accepted packets from firewalls?
You can be highly selective about which events are forwarded to the Splunk indexer. One conceptual model just to illustrate the point is given below:
Threat Hunting
Threat Hunting is kind of the sexy offering for the world of defence. Offence has had more than its fair share of glamour offerings over the years. Now its defence’s turn. Or is it? I mean i get it. It’s a good thing to put on your profile, and in some cases there are dramatic lines such as “be the hunter or the hunted”.
However, a rational view of “hunting” is that it requires LOTS of resources and LOTS of skill – two commodities that are very scarce. Threat hunting in most cases is the worst kind of resources sink hole. If you take vulnerability management (TVM) and the kind of basic detection discussed thus far in this article, you have a defence capability that in most cases fits the risk management needs of the organisation. So then there’s two questions to ask:
How much does threat hunting offer on top of a suitably configured logging and TVM capability? Not much in the best of cases. Especially with credentialed scanning with TVM – there is very little of your attack surface that you cannot cover.
How much does threat hunting offer in isolation (i.e. threat hunting with no TVM or logging)? This is the worst case scenario that will end up getting us all fired in security. Don’t do it!!! Just don’t. You will be wide open to attack. This is similar to a TVM program that consists only of one-week penetration tests every 6 months.
Threat Intelligence (TI)
Ok so here’s a funny story. At a trading house client here in London around 2016: they were paying a large yellow vendor lots of fazools every month for “threat intelligence”. I couldn’t help but notice a similarity in the output displayed in the portal as compared with what i had seen from the client’s Alienvault. There is a good reason for this: it WAS Alienvault. The feeds were coming from switches and firewalls inside the client network, and clearly $VENDOR was using Alienvault also. So they were paying heaps to see a duplication of the data they already had in their own Alienvault.
The aforementioned is an extremely bad case of course. The worst of the worst. But can you expect more value from other threat intelligence feeds? Well…remember what i was saying about the value of an effective TVM and detection program? Ok I’ll summarise the two main problems with TI:
You can really achieve LOTS in defence with a good credentialed TVM program plus even a half-decent logging program. I speak as someone who has lots of experience in unrestricted penetration testing – believe me you are well covered with a good TVM and detection SecOps function. You don’t need to be looking at threats apart from a few caveats…see later.
TI from commercial feeds isn’t about your network. Its about the whole planet. Its like picking up a newspaper to find out what’s happening in the world, and seeing on the front cover that a butterfly in China has flapped its wings recently.
Where TI can be useful – macro developments and sector-specific developments. For example, a new approach to Phishing, or a new class of vulnerability with software that you host, or if you’re in the public sector and your friendly national spy agency has picked up on hostile intentions towards you. But i don’t want to know that a new malware payload has been doing the rounds. In the time taken to read the briefing, 2000 new payloads have been released to the wild.
Summary
Start out with a design process that takes input feeds from compliance and risk (perhaps threat modelling), use the resulting requirements to drive the LLD, which may or may not result in a decision to procure tech that meets the requirements of the LLD.
An effective logging capability can only be designed with intimate knowledge of the estate – databases, crown jewels, data flows – for each application. Without such knowledge, it isn’t possible to build even a barely useful logging capability. Call out your generic and custom use cases in your LLD, independent of technology.
Get your basic alerting first, correlation can come later, if ever.
Outsourcing is a waste of resources for second level SecOps.
With SaaS, your SIEM itself is dangerously exposed, and you have no control over what is logged from SaaS log sources.
You are not mandated to get a COTS. Think about what it is that you want to achieve. It could be that open source tools across the board work for you.
Splunk really is the Lamborghini of SIEMs and the “expensive” tag is unjustified. If you carefully design custom and generic use cases, and remove everything else from indexing, you suddenly don’t have such an expensive logger. You can also aggregate everything in a Syslog pool before it hits Splunk indexers, and be more selective about what gets forwarded.
I speak as someone with lots of experience in unrestricted penetration testing: Threat Hunting and Threat Intelligence aren’t worth the effort in most cases.