Loading...
Loading...
Proactive threat hunting is a hypothesis-led search of your telemetry for behavior no detection rule was built to catch, assuming an attacker may already be inside. Coverage is thinner than heatmaps show: in September 2026, MITRE found that 144 public Sigma rules for Windows Security logs mapped to 84 techniques, but only 26 of those techniques had any implementation-level coverage.
Count the green cells on your MITRE ATT&CK coverage heatmap. Each one means at least one rule maps to that technique. Which versions of the technique that rule can see is a separate question, and an attacker only needs one it can't.
On October 8, 2026, the FBI, CISA, the NSA, and agencies in six other countries spelled out one of those versions in joint advisory AA26-281A. Chinese government-linked actors, enabled by a company called Integrity Technology Group, installed SoftEther VPN on victim machines to keep remote access. They often named the installers conhost.exe or dllhost.exe and set the client to reconnect at every startup. Endpoint detection, the advisory notes, "is less likely to flag SoftEther because it is a legitimate VPN software."
Proactive threat hunting is a search for attacker activity that starts from a hypothesis instead of an alert, on the assumption that an intruder may already be inside where your automated detection can't see them. Reactive hunting starts from something that already happened. An alert fires, an incident opens, or a feed publishes an indicator.
The SoftEther advisory asks readers to "hunt for potential compromises from this activity," which is reactive hunting with the answer key attached.
The proactive version asks the same question before anyone publishes, while the attacker is still undiscovered. Mandiant's M-Trends 2026 tells organizations to "implement a routine of proactive threat hunting," because reactive alerts alone aren't enough to catch sophisticated, evasive attackers. It also put median dwell time at 14 days in 2025, up from 11 in 2024.
Half the intrusions Mandiant investigated in 2025 ran two weeks or longer before anyone found them. A proactive hunt doesn't wait for an alert or an advisory, and it typically runs in four steps:
A green ATT&CK heatmap misleads because it marks a technique covered when any rule maps to it, while attackers use specific implementations of that technique the rule may never see. MITRE's Center for Threat-Informed Defense (CTID) defines an implementation as "a behaviorally-distinct way of executing a technique, defined by execution path and required system interactions." The SoftEther installs map to T1133, External Remote Services. A rule written for some other remote-access path can turn that cell green without ever seeing a SoftEther client.
In the September 10 post introducing CTID's Detection Coverage Calculator, part of Summiting the Pyramid 5.0, Antonia Feffer wrote that "two organizations can show the same technique as green on an ATT&CK heatmap while having dramatically different capabilities behind that box."
CTID's Windows Security study scored 144 public rules written in Sigma, the open format for sharing SIEM detections, that map to 84 techniques. Implementation coverage came to 26.8%, and only 26 of the 84 techniques had any implementation-level coverage at all.
In November 2026, Bryan Hsu and Feffer ran 773 Windows Sigma rules against the Volt Typhoon playbook. Of the 42 techniques those rules mapped to, 95.2% had a covered implementation. Of the 168 implementations behind them, 78.0% did. In what CTID calls an exploratory join, only 17.3% of those implementations had a rule scoring 3 or higher on both robustness (how hard a rule is to evade) and precision.
On a heatmap, those 42 techniques would look solid green. Zoom to implementations and more than one in five has no rule that can see it. Ask for a rule that is both hard to evade and precise, and fewer than one in five qualifies.
CTID is careful about what this proves. The results reflect "modeled coverage from a selected set of Windows Sigma analytics," and CTID says they don't predict whether a real network would catch Volt Typhoon. Your own rules may well beat public Sigma, but an average flatters you either way, because "an adversary doesn't need to defeat every detection associated with a technique; it only needs to defeat one."
Signature-based detection matches artifacts an attacker brings and can change, such as hashes, IP addresses, domains, file names, and command strings. Behavior-based detection, sometimes called behavioral detection, matches what the attacker has to do on your systems to reach a goal. The main difference is what the attacker must change to get past each one. A new hash or file name beats a signature, and a behavior rule holds until the attacker switches implementation or technique.
A rule keyed to the file name on disk misses an installer saved as conhost.exe. A rule on what the client has to do, connecting out to an attacker-controlled VPN server every time the machine starts, doesn't care what the file is called.
| Feature | Signature-based detection | Behavior-based detection | Proactive threat hunting |
|---|---|---|---|
| Matches | Known artifacts such as hashes, IPs, domains, and file names | Actions a technique requires on the system | A hypothesis about behavior, tested over weeks or months of telemetry |
| Catches | Threats whose artifacts someone has already seen | New tools that still perform a covered behavior | Activity no rule can decide alone, such as misuse of legitimate tools |
| Runs | On every event | On every event | On a cadence or a trigger |
| Evaded by | Changing the artifact | Switching implementation or technique | Telemetry nobody collected |
| Needs updating when | A new artifact appears | The technique gains a new implementation | Telemetry or context sources change |
| False positives | Few on exact matches, though stale indicators need vetting before blocking | Higher where administrators perform the same actions | Weighed against context before a verdict |
| Decides intent from | What reaches it at run time, usually the event | What reaches it at run time, usually the event | The event plus context (owner, baseline, change record) |
| Output | Alert | Alert | Verdict, rule candidate, or telemetry gap |
Signatures still earn a place as sweeps and fast blocks.
CTID's Windows study agrees that "specific indicators still have value, particularly for threat hunting and rapid response." They also age. The SoftEther advisory warns that several of its own indicators "date back to as early as 2016" and asks defenders to vet them before blocking anything.
The Pyramid of Pain is David Bianco's 2013 model that ranks indicators by how much it costs an attacker to lose them. Hash values, added in a 2014 revision, sit at the base. Tactics, techniques, and procedures sit at the top, where detecting and responding forces adversaries to "learn new behaviors." CTID's Summiting the Pyramid turns that ranking into a score for each observable in a rule:
A rule scores at its weakest required observable. CTID's worked example, the public Sigma rule "CobaltStrike Named Pipe Patterns," lands at Level 1 because "these pipe names can be easily changed by the adversary."
In the SoftEther advisory, the installer's file name is Level 1. The same actors also ran a tool that bound to a domain controller over RPC and pulled data through the Directory Replication Service. That's DCSync, and CTID uses its replication calls as a Level 5 example, since DCSync implementations using that mechanism have to make them. The advisory's own mitigations include monitoring for unexpected Active Directory replication.
A detection rule can't decide an attack whose malicious and legitimate versions produce the same events, unless the fact that separates them reaches the rule at run time. CTID calls these ambiguous techniques, because their observables can't establish intent with confidence. The deciding fact is usually the account owner, a change approval, or what that host normally does.
Hunters have known this for years. MITRE's TTP-Based Hunting report (2019) found that system administrator activity "is one of the most common sources of false positive events." Six years later, CISA's Scattered Spider advisory (AA23-320A), updated in July 2025, still tells defenders that the legitimate remote access tools the group used, such as TeamViewer, AnyDesk, and Tailscale, are not on their own a sign of malicious activity.
The SoftEther advisory's own mitigations include a check that depends entirely on context: watch for high upload volumes from workstations "that usually have low volumes of uploads compared to downloads."
No rule can use that check until someone has measured, host by host, which of your workstations normally upload little. That measurement is the hunt.
Proactive threat hunting should start where your rule set is structurally weakest: technique implementations no rule covers, ambiguous techniques whose intent depends on context, and rules that score Level 1 or 2 on Summiting the Pyramid. Each gives a hunter a first query.
If you only have time for one, take the ambiguous techniques, since the other two can at least end in a better rule.
Pick a technique that shows green, such as T1133, the one the SoftEther advisory names. List its implementations and match each to a rule. ATT&CK's Detection Strategies, added in v18 in October 2025, help build the list. Each implementation with no analytic behind it is a threat hunting hypothesis you can test this week, and for T1133 the first is a legitimate VPN client, under a borrowed Windows name, that dials out at every boot. If you're unsure where else to begin, CTID's November 2026 Volt Typhoon study found that Discovery accounted for 30 of the 37 uncovered implementations.
Where intent depends on context, run a baseline hunt first. Splunk's PEAK framework (Prepare, Execute, and Act with Knowledge) describes it as establishing "normal" behavior and searching for deviations, and Bianco, who later co-created PEAK, warns that "just because something is abnormal doesn't mean it's malicious." For the advisory's upload check, the baseline is each workstation's ratio of uploads to downloads. Keep it whatever the hunt finds, since a future rule on upload volume will need it.
List the rules that would score Level 1 or 2 and hunt the behavior they were written for. A rule that matches the SoftEther installer by name sits at Level 1, and in the published case the actors often saved it as conhost.exe or dllhost.exe. Hunt the behavior behind it instead, and you'll find out whether a stronger rule is even possible. Rules also decay as your environment changes, and keeping them current is detection engineering work.
Sometimes the hunt stalls at the first step. MITRE's TTP-Based Hunting method expects this and makes "identify and mitigate collection gaps" a step of every hunt. If nobody collects per-host upload volume, the hunt stops there and reports the missing log source.
Summiting the Pyramid, a MITRE CTID project, scores how hard a rule built from a hunt finding would be to evade. At Level 3 or higher, an attacker generally has to change approach or target to get past it. At Level 1 or 2, the finding is better kept as a sweep or a dashboard.
In its Volt Typhoon study, CTID notes that its score-of-3 threshold "is not a pass/fail standard," so treat it as a working line. PEAK supplies the ladder of outputs, from reports and dashboards up to analytics in code and then signatures and rules, with the advice to "choose the highest level that you can feasibly implement." The level score tells you how high is feasible.
For an ambiguous technique, the score isn't enough. The deciding context, such as an asset owner or a change ticket, has to be available to the detection too, or the finding stays a scheduled hunt and the missing context source becomes a gap to close.
Run SoftEther through that test. "SoftEther ran" describes a tool the attacker chose and can swap. Matched on the binary's metadata, that's Level 2, and matched on the file name, it drops to Level 1. "A workstation that rarely uploads started sending large volumes out" holds whatever tool sends the data, so it's the better candidate, provided each host's upload baseline reaches the detection. The project's robustness levels give no example for a volume baseline, so treat that call as judgment. The rule won't catch everything either. The same advisory describes a tool built to pull mail straight from Outlook 365 accounts, a path that may never involve one of your workstations.
AI helps with the parts of proactive threat hunting that don't scale by hand, which are testing many hypotheses against months of telemetry and ruling out benign explanations before anything is called a threat. With AI threat hunting, deciding what to hunt, and which rules ship, should still stay with your team.
Simbian's AI Threat Hunt Agent is built on that split. It plays what our lead threat hunting engineer calls the elimination game, testing the benign explanations first and calling a threat only when they fail. In context-aware mode it draws on the Context Lake™ to rule out activity your organization already knows about, such as an executive logging in from a business trip.
But a note that explains away a real executive's login can explain away an intruder using valid access just as easily. SoftEther is that trap. The reason endpoint detection is less likely to flag it, that it's legitimate VPN software, is the same reason an approved-software note would wave it through. Log-first mode sets the context aside and reads raw logs for questions like that one, where trusting the context is the risk.
When the activity was legitimate this time but followed an attack-shaped pattern no rule covers, the Agent's verdict is Detection Opportunity. The AI Detection Engineering Agent can draft that candidate rule, and nothing deploys until a person approves it. Anything the hunt couldn't check, such as a log source it couldn't reach, shows up in its report as a gap, flagged for whether it affected the verdict.
Your heatmap will keep showing T1133 green, whether or not any rule can see a renamed SoftEther client.
Pick one green cell this week, list its implementations, and find out which ones your rules can see. To watch the AI Threat Hunt Agent work through your own telemetry, book a demo.
Q: Does proactive threat hunting replace detection rules? No. Detection rules check every event continuously, and proactive threat hunting tests what those rules can't decide on their own. A good hunt ends by improving the rule set, through a new rule, a stronger observable for an existing one, or a telemetry gap closed.
Q: Can signature-based detection detect zero-day attacks? Usually not. A signature matches an artifact someone has already seen, so a new exploit or a renamed tool passes until the signature is written. MITRE's TTP-Based Hunting report says indicator sweeping "fails to identify novel or changing threats that don't match known indicators." Behavior-based rules and hunts look instead at what the attacker has to do after the exploit lands.
Q: Can behavior-based detection replace signature-based detection? No. Behavior-based rules hold up when attackers rename or rebuild their tools, but signatures still work as sweeps and fast blocks for indicators someone has already seen. Indicators age, though, and CISA's SoftEther advisory (AA26-281A) asks defenders to vet its own, some dating to 2016, before blocking on them.
Q: What does anomaly-based detection look for? Anomaly-based detection looks for deviation from your own baseline, such as a workstation that suddenly uploads far more than it usually does, whether or not the deviation maps to a known technique. Behavior-based detection looks for actions a technique requires, whether or not they're unusual for you, and signature-based detection looks only for artifacts someone has already seen. Baseline hunts use the anomaly approach to find the context a behavior rule needs.
Q: What are the three main types of threat detection? Signature-based, anomaly-based, and TTP-based (behavior-based) detection. Signatures match known artifacts such as hashes and file names, anomaly detection flags deviation from your own baseline, and TTP-based detection matches what an attacker has to do to carry out a technique. MITRE's TTP-Based Hunting report calls them "complementary approaches," and a hunt program draws on all three.
Q: What is Summiting the Pyramid? Summiting the Pyramid is a MITRE Center for Threat-Informed Defense (CTID) project that scores how hard a detection rule is to evade. Building on David Bianco's Pyramid of Pain, it rates each observable on five levels, from ephemeral values like hashes to behaviors every implementation of a technique must show. Version 5.0, released in September 2026, added a detection coverage calculator.
Q: How often should you run proactive threat hunts? Run them on a fixed cadence for your highest-risk techniques, and on a trigger when a new advisory, log source, or rule change lands. A hunt that becomes a rule no longer needs repeating by hand, so treat any hunt you keep rerunning as a rule you haven't written yet.