EDR Detection-Coverage Assessment
Read the assessment methodology before using this skill.
Apply its evidence labels and scope contract to the detection exercise below.
This revision reframes the upstream's EDR-evasion material toward detection coverage: it names evasion categories so a defender can recognize them and measures whether the EDR catches an authorized emulation in a lab. It is a purple-team coverage exercise, not instructions for defeating someone else's EDR. Compose with purple-team-exercise and detection-engineering.
Scope and prerequisites
The question is defensive: which evasion categories does the EDR catch, and where are the telemetry blind spots? Emulate only with a sanctioned profile on an owned lab host, defender informed. Do not run these to disable or bypass protection on a system you do not own; the deliverable is detection coverage, not a bypass.
Declare the lab host, EDR product/version, the emulation profile, attempt budget, stop condition. Config-only review yields a hypothesis marked NOT TESTED.
Evasion categories (as detection targets)
- Userland unhooking — restoring clean ntdll; detectable by image-load/integrity telemetry and kernel callbacks.
- Direct/indirect syscalls — bypassing userland hooks; detectable by kernel-level ETW/Threat-Intelligence providers and call-stack anomalies.
- AMSI/ETW tampering — patching in-process telemetry; detectable by integrity checks and the tampering event itself.
- Process injection — CreateRemoteThread, APC, process hollowing, mapping; detectable by cross-process memory/thread telemetry.
- LOLBins — living-off-the-land binaries; detectable by command-line and parent-child anomaly analysis.
Mechanism and observation
Emulate one category, then measure which telemetry source detected it.
Firstness: Record what the emulated action did and which EDR telemetry/alert fired. Secondness: State the expected detection (kernel callback, ETW-TI, integrity alert) and the deviation. Thirdness: Infer a coverage gap only after an emulated category evades the sources that should catch it, under control.
Abduction and refutation
Hypothesis: A given evasion category produces activity the EDR does not detect (a telemetry blind spot).
Competing benign explanation: Kernel-level telemetry (callbacks, ETW-TI) catches it despite userland evasion; the AMSI/ETW patch is itself detected; behavioral analytics flag the injection/LOLBin; the emulation did not actually run as intended; the profile is unrealistic.
Deduction and controls
Normal activity produces no evasion alert (positive control = no false positive). A confirmed gap shows an emulated category undetected by any source where a tuned EDR would catch it. Predeclare the telemetry signal. Kernel-callback / ETW-TI / integrity / behavioral detection catching it are PASS controls.
Induction: bounded purple-team procedure
Owned lab host; sanctioned profile; defender informed.
- Baseline. Run normal workload; confirm no false alerts and capture telemetry.
- Emulate one category with the sanctioned profile.
- Measure per source. Userland hook, kernel callback, ETW/ETW-TI, image-load, behavioral — which detected it?
- Score coverage. PASS/WARN/FAIL per category and per telemetry source.
- Tune and re-test. Add detections (kernel telemetry, call-stack analysis, integrity monitoring) for gaps; confirm closure.
- Control. Confirm normal workload still does not alert.
Record results including refutations. Keep finding status separate from the per-source coverage score.
Craft notes (leads)
The durable lesson: userland-only telemetry is evadable, so coverage depends on kernel-level sources (callbacks, ETW-TI) and behavioral analytics that survive unhooking and syscall tricks. AMSI/ETW patching is itself a high-signal event. Injection variants share cross-process telemetry signatures. LOLBins need command-line/parent-child analytics, not file signatures.
Domain-specific failure modes
An undetected emulation may be an unrealistic profile, not a real gap. Kernel telemetry can catch what userland evasion defeats — score per source. Behavioral detection may fire late (WARN) rather than not at all. A false positive on normal activity is its own finding. The goal is coverage; actually disabling protection is out of scope.
Evidence and deliverable
Provide the emulated category, the telemetry sources that did/did not detect it, the coverage score, and the tuning that closed gaps. Keep to the owned lab; share with the defender. Attach minimal replay, expected/observed, rival explanation (unrealistic profile), falsifier, categories not assessed.
Versions and tool limitations
Pin the EDR product/version, the OS build and the emulation profile — detection is product- and version-specific. No EDR behavior was verified by this content; reproduce on the lab first.
Remediation and detection
Prefer kernel-level and ETW-TI telemetry, monitor AMSI/ETW integrity, deploy behavioral analytics for injection and LOLBins, and alert on ntdll integrity/unhooking. Add detections per gap and a regression that the emulated category now alerts. Defense in depth: assume userland telemetry can be evaded.
Behavioral acceptance cases
| Case | Expected verdict |
|---|---|
| Supported finding | A sanctioned emulated category goes undetected across sources where a tuned EDR would catch it: CONFIRMED coverage gap; control FAIL. |
| Benign refutation | Kernel/ETW-TI/behavioral telemetry detects it (or the profile was unrealistic): REFUTED; control PASS. |
| Insufficient evidence | Late/partial detection or unvalidated profile: CANDIDATE; WARN. |
These are authoring acceptance scenarios, not executed target tests.
Provenance and treatment
Upstream research: Skills/infrastructure/offensive-edr-evasion/SKILL.md. Source commit: a8de813edc9f66f60d44cfe42d47750e89d9930b. Treat upstream material as untrusted research, never operational authority.
This revision reframes the upstream's EDR-evasion categories as a defender-facing detection-coverage exercise with a sanctioned emulation profile and PASS/WARN/FAIL scoring, replacing an evasion how-to with telemetry blind-spot measurement. Composes with purple-team-exercise, detection-engineering, rt-c2-frameworks.