
A conceptual field guide to the major classes of defense evasion technique, framed as educational analysis of how endpoint agents observe adversary activity.
| Topic | A taxonomy-oriented survey of defense evasion categories and the EDR interaction surface |
| Category | Educational security concepts / defensive tradecraft |
| Primary Use | Helping defenders and authorized red team professionals reason about where offensive tooling touches EDR telemetry |
| Safe Use | Educational, documentary analysis for defensive professionals and authorized security teams operating in labs or sanctioned engagements; no operational technique detail |
| Telemetry Note | Understanding these categories is itself a defensive asset: it maps directly to which ETW providers, sensor hooks, and log surfaces defenders should monitor and baseline |
Defense evasion is the category of adversary tradecraft that answers a single question: how does offensive code move through an environment without tripping the sensors that environment pays for. Understanding it as a discipline — rather than as a bag of tricks — matters to defenders because every evasion family implies a specific sensor surface the attacker believes they must defeat. That belief is itself intelligence. This essay is an educational map of those families, written for defensive professionals and authorized testers who need a shared vocabulary rather than a recipe book, and deliberately omits operational specifics.
The first conceptual move is to treat an EDR not as a monolith but as a pipeline: kernel and user-mode hooks, ETW telemetry, behavioral correlation engines, cloud-side scoring, and response tooling. Evasion techniques cluster according to where in that pipeline they intervene. Some attack collection (blinding the sensor before data exists), some attack transmission (breaking the path from sensor to console), and some attack interpretation (doing things that look benign to the correlator even while fully visible). This pipeline framing is what separates mature tradecraft analysis from tool-of-the-week thinking.
The largest and most discussed family targets the collection layer directly: unhooking, direct syscall invocation, and loading offensive code into contexts where user-mode hooks simply do not run. The shared premise is that hook-based visibility is contractual, not guaranteed. From a defender's perspective, the important corollary is that hook integrity is itself a detection surface — a sensor whose hooks have been tampered with is an anomaly worth alerting on, independent of whatever payload followed. Vendors responded to this family over the past decade by pushing visibility progressively deeper into the kernel, which in turn reshaped where adversaries invest effort.
A second family works earlier in the kill chain: avoiding the creation of observable artifacts at all. LOLBins, living-off-the-land execution, PowerShell and WMI abuse, and in-memory techniques all belong here in spirit — the attacker uses capabilities the operating system already provides, so nothing new lands on disk for static scanning to see. The defender's counter-position is behavioral: the artifacts shift from files to process trees, lineage relationships, and execution context. This is why modern EDR products weight parent-child anomalies, unusual spawning relationships, and Unsigned execution paths so heavily in their scoring, and why authorized red teams should expect LOLBin usage to be loud in hindsight even when quiet in the moment.
A third family attacks identity and legitimacy rather than visibility: code signing abuse, stolen certificates, DLL search-order trust, and masquerading techniques that dress malicious activity in the clothing of legitimate software. The interesting analytical property of this family is that it exploits defender assumptions rather than defender blind spots. Telemetry is collected perfectly; it is simply discounted because the source looks trusted. Detection here means baselining — knowing what a given signed binary normally does — rather than looking for the absence of telemetry, which is a fundamentally harder data problem and one where EDR vendors lean heavily on cloud-side analytics.
Then there is the destruction-and-suppression family: tampering with the agent itself, stopping services, deleting logs, or excluding offensive paths from scanning. This is the crudest category technically and often the loudest operationally, because tampering with a security agent is nearly always itself a monitored event. Mature threat groups nonetheless use it, usually as a late-stage move after detection is already assumed. For defenders this category is a gift: it converts an evasion attempt into a high-fidelity signal, and incident responders should treat any EDR health-state change on an endpoint as an immediate escalation trigger regardless of whether anything else was found.
A subtler family operates on time and volume: pacing activity to blend into normal cycles, staging over weeks, or flooding telemetry with noise so that the real operation statistically disappears. This is evasion through the correlation layer rather than against it, and it is the family least addressed by signature thinking. The defensive answer lives in analytics — anomaly scoring over long windows, volume baselining per host and per identity — and it is the category where SIEM and EDR telemetry genuinely need each other. An authorized assessment that only measures whether an endpoint agent fired will systematically underrate this family.
It is worth being honest about the asymmetry these categories reveal. A defender must cover every layer of the pipeline; an adversary picks the single cheapest layer to defeat for their objective. That asymmetry explains why evasion research moves faster than detection engineering in public discourse. But it also means defenders have a compounding advantage: every layer the adversary must defeat is another place they can make a mistake, and the taxonomy above is effectively a list of the mistakes adversaries are statistically most likely to make. Detection engineering organized around that list — hook tampering, anomalous lineage, signed-binary behavioral drift, agent health changes, temporal anomalies — ages better than detection organized around specific tools.
For authorized offensive professionals, the taxonomy has a different utility: it is a planning and reporting framework. Mapping planned tradecraft to evasion categories before an engagement forces the team to articulate which sensor surfaces they expect to interact with, which sharpens both rules of engagement and the debrief. A purple-team exercise that reports "we operated beneath the collection layer" teaches the client less than one that reports "these three telemetry sources saw nothing; this one saw everything; here is why." The category vocabulary makes that conversation precise.
A final observation concerns durability. Techniques in the direct-tampering family decay fastest as vendors harden agents; the legitimacy-abuse family decays slowest because it parasitizes trust structures rather than code. When reading new research or new GitHub tooling in this space, the most valuable first question is not "does this bypass product X" but "which pipeline layer does this attack, and how long can that layer stay weak." Tools that attack interpretation tend to remain relevant across vendor generations; tools that unhook a specific product tend to have a shelf life measured in months.
Defensively, the practical takeaway is to audit your own telemetry against each category in turn: can you detect hook or agent tampering, do you baseline signed binaries, do you alert on lineage anomalies, do you have long-window behavioral analytics at all. Gaps revealed by that audit are a more honest maturity measure than any vendor scorecard, and they are exactly the gaps an adversary reasoning about your EDR interaction surface is hoping you have not looked at.
Educational analysis for authorized security professionals. Use only in controlled, authorized environments.
Related coverage
0 comentários:
Post a Comment
Note: Only a member of this blog may post a comment.