Think Like a Detection Engineer, Pt. 1: Logging
Instrumenting our security data
Detection Engineering sits at the intersection of InfoSec, Cloud Infrastructure, DevOps, and Software Development. In this post, I’ll step through the thought process of a Detection Engineer in the context of collecting security data.
What does a Detection Engineer do?
Detection Engineers build and deploy systems that validate security controls and detect suspicious behaviors with code. Our goal is to protect the “Crown Jewels” and prevent incidents in the organizations we serve.
Getting started as a Detection Engineer involves mapping and classifying systems (and data) by importance. Then, with this understanding, “detections” are created that flag varying degrees of behaviors by risk level.
You can’t detect what you can’t see. So let’s start with getting data.
The layers of Security Data
What is “security data”?
This refers to audit logs of past behaviors, used for security monitoring. They exist in almost all systems we use every day:
$ ls /var/log alternatives.log apache2 auth.log cloud-init-output.log dist-upgrade journal landscape lxd osquery syslog td-agent wtmp amazon apt btmp cloud-init.log dpkg.log kern.log lastlog mail.log suricata tallylog unattended-upgrades $ tail /var/log/auth.log <38>1 2022-07-18T00:41:20.962241+00:00 ip-172-31-29-253 sshd 27835 - - Invalid user sftpuser from 126.96.36.199 port 44042 <38>1 2022-07-18T00:41:36.574327+00:00 ip-172-31-29-253 sshd 27835 - - Connection closed by invalid user sftpuser 188.8.131.52 port 44042 [preauth] <38>1 2022-07-18T00:59:19.284071+00:00 ip-172-31-29-253 sshd 27915 - - Invalid user dev from 184.108.40.206 port 54074 <38>1 2022-07-18T00:59:19.377547+00:00 ip-172-31-29-253 sshd 27915 - - Received disconnect from 220.127.116.11 port 54074:11: Bye Bye [preauth] <38>1 2022-07-18T00:59:19.377731+00:00 ip-172-31-29-253 sshd 27915 - - Disconnected from invalid user dev 18.104.22.168 port 54074 [preauth]
Audit logs exist in several places in our environment, which can be broken down this way:
Infrastructure (think AWS, GCP, Azure, and their services)
Host (laptops, VMs, baremetal systems)
Network (netflow, IDS/IPS)
Application (internet-facing applications or SaaS)
Database (DML/DDL transactions)
Each layer provides context to the overall picture, and in fact, the same behavior can span multiple layers at once.
As an example, let’s look at four different logs from one SSH session (joined by time and source port)
Throughout each log, we learned different things about that session. Are all of them necessary? Definitely not. But it helps fill in information gaps along with creating redundancy in our pipeline.
Tradeoffs of collecting logs
Collecting security logs comes with a series of tradeoffs:
Signal: How valuable is the data on its own for detection? Does the log have enough context for responders to understand what had happened?
Latency: How long does it take to get the data into the SIEM? Depending on the service, this could range from <1 minute to more than 30 minutes. We want to detect as fast as possible.
Cost: How expensive is it to get and retain the data? Cost can also reference performance hits on production systems.
Ideally, we optimize for relevant, high-signal logs. Less is more here, especially when it comes to fast response. The more data, the slower it is to enumerate it.
For instance, I would optimize for Osquery Logs in the above example because they provide the richest data on the overall connection and enable me to analyze them in isolation with higher confidence.
Consolidating data into a single place
Before we can create detections, data has to be centralized into the SIEM.
SIEMs are used as the brain of your security monitoring pipeline, where all data and detection logic lives.
But getting them there is non-trivial, especially in large organizations with complex cloud infrastructure. Logs are also in different formats and must be normalized in order to adequately search and analyze them.
OSS tools like Fluentd and Logstash can be great tools for getting host-based logs into cloud storage, like S3, which SIEMs can consume from. They speak multiple protocols, like Syslog and TCP, allowing for more audit log variety.
Again, always prioritize logs that can give you the best signal at the lowest latency/cost.
Having confidence in the pipeline
Operational monitoring is the last, critical part of Security Data collection.
Security Monitoring is a continuous stream of logs, and if that stops for some reason, it creates a blindspot for proactive detection and reactive response.
Metrics and alarms should always be configured for this purpose, and failover/redundancy should also ideally exist.
Getting logs is the first step
Getting high-quality data for security monitoring is a critical first step for any detection engineer.
It takes time to instrument all of the necessary systems, normalize that data, and make it operational. The good news is that these are typically one-time investments that payoff for the longer term of the security team and the overall posture of your organization.
With this data, we can create Detections to flag attacker behaviors and have confidence that our data will continually flow (and be on time).
Next time - we will talk about crafting detections.
Thanks for reading Detection at Scale! Subscribe for free to receive new posts and support my work.