Ransomware operators now routinely switch off endpoint detection and response (EDR) before they encrypt, and the most reliable way they do it is by loading a legitimately signed but vulnerable kernel driver. This technique, bring your own vulnerable driver (BYOVD), gives an attacker who already holds local administrator rights a path into the Windows kernel, where security agents can be terminated or their callbacks removed. Packaged "EDR killer" tools such as AuKill, EDRKillShifter and the driver mode in DragonForce's encryptor have turned it into a commodity sold to affiliates. This guide explains the mechanism at a defensive level, where it appears, what to log, and how the Microsoft vulnerable driver blocklist, memory integrity (HVCI) and App Control for Business close the gap. It is for endpoint engineers, detection engineers and incident responders, and goes deeper on drivers than the broader EDR Evasion and Defence Guide.
Key takeaways
- BYOVD is a post-compromise step, not an entry point. The attacker needs administrator rights to install a driver, so local admin control and privileged identity hygiene are the first line of defence.
- Signed does not mean safe. Windows accepts the driver because a real vendor signed it; the flaw is in what the driver lets any caller do once loaded.
- A small set of drivers does most of the damage. ESET found EDR killers rely on a limited pool of public vulnerable drivers and proof-of-concept code, which makes blocklisting effective.
- The blocklist only works if it is enforced and current. Microsoft's in-box list is on by default on Windows 11 2022 update devices with memory integrity, but the downloadable list is more complete and already-running drivers are only blocked after a reboot.
- Detect the driver load, not the aftermath. A new kernel service, a driver load from a user-writable path or a known-bad hash is the last reliable signal before the EDR goes quiet.
How the attack works
Windows requires kernel drivers to be signed. Over many years, hardware and software vendors have shipped signed drivers that expose dangerous functions to user mode: reading and writing arbitrary kernel memory, terminating any process, or mapping physical memory. These drivers remain validly signed after the flaw is published, so an attacker can bring an old copy onto a host and load it.
- Prerequisite: the attacker already has administrator or SYSTEM rights, typically from stolen credentials or an earlier exploit (MITRE ATT&CK T1068 Exploitation for Privilege Escalation describes adversaries bringing a signed vulnerable driver onto a compromised machine for kernel execution).
- Staging: the driver file is dropped to disk, often under a renamed or generic filename.
- Loading: the driver is registered as a kernel service (T1543.003 Windows Service) and started.
- Abuse: the EDR killer sends requests to the driver to terminate security processes, remove kernel callbacks or tamper with protected processes (T1685 Disable or Modify Tools). Some tools loop, killing the agent again whenever it restarts.
- Follow-on: with the sensor blind, operators run credential theft, stop services (T1489 Service Stop) and deploy ransomware.
A related variant is the malicious signed driver: attacker-written code signed with a stolen or fraudulently obtained certificate (T1588.003 Code Signing Certificates, T1553.002 Code Signing). The NVIDIA certificates leaked in 2022 are one example: the NVIDIA code-signing certificate theft led to malware, including a Windows driver, being signed with expired NVIDIA certificates, which Windows driver signing rules could still accept. Both variants end in the same place: attacker-controlled code in ring 0.
| Variant | What is abused | Typical control that stops it |
|---|---|---|
| Vulnerable legitimate driver | Exposed IOCTLs in an old vendor driver (for example Process Explorer, Dell dbutil, GIGABYTE GDrv, Intel iqvw64) | Vulnerable driver blocklist, App Control deny rules, hash detection |
| Malicious driver with stolen certificate | Code-signing trust in a compromised certificate | Certificate-based deny rules, blocklist updates, revocation |
| Rootkit after kernel access | Kernel hooks to hide files, processes or network activity (T1014 Rootkit) | Memory integrity, Secure Boot, offline forensics |
Where it shows up
EDR killing has moved from a rare skill to a standard ransomware stage. The tools below come from vendor research and government advisories that defenders can read in full.
| Tool or actor | Driver abused | Reported use | Source |
|---|---|---|---|
| AuKill | Outdated Process Explorer driver v16.32, dropped as PROCEXP.SYS | Medusa Locker and LockBit intrusions, early 2023; earliest sample November 2022 | Sophos |
| EDRKillShifter | Drivers including rentdrv2.sys and TFSysMon | Built by RansomHub, introduced 8 May 2024 and offered to affiliates; later used by an affiliate ESET calls QuadSwitcher across Play, Medusa and BianLian | ESET |
| Medusa ransomware | Vulnerable or signed drivers | Used to kill or delete EDR | CISA AA25-071A |
| DragonForce encryptor | truesight.sys, rentdrv2.sys (optional mode, drivers must already be present) | Process termination before encryption | Acronis |
Several drivers abused this way are in CISA's Known Exploited Vulnerabilities catalogue, which helps prioritise blocking and inventory.
| CVE | Driver | KEV date added | Known ransomware use (KEV) |
|---|---|---|---|
| CVE-2015-2291 | Intel Ethernet Diagnostics Driver (iqvw64.sys) | 10 February 2023 | Known |
| CVE-2018-19320 | GIGABYTE GDrv | 24 October 2022 | Known |
| CVE-2018-19321 | GIGABYTE GDrv | 24 October 2022 | Known |
| CVE-2021-21551 | Dell dbutil driver | 31 March 2022 | Unknown |
The community LOLDrivers project, maintained by MagicSword with GitHub contributions, listed 706 entries (581 vulnerable, 125 malicious) covering 2,417 samples when checked, and publishes Sigma, YARA, Sysmon and App Control rules. It also records around 19.8% of samples as loading even with HVCI enabled, a reminder that memory integrity is not a complete answer on its own. Driver ownership is covered in the RansomHub Threat Actor Profile and Medusa Ransomware Threat Actor Profile.
Impact
- Loss of visibility: once the agent is killed, the SOC loses process, network and file telemetry from that host, often for exactly the window in which credentials are dumped and ransomware runs.
- Silent failure: some consoles show the host as merely offline or unhealthy. Without a specific "sensor stopped unexpectedly" alert, the gap looks like a network issue.
- Kernel-level persistence: where a malicious driver is loaded, rootkit functions can hide artefacts and survive reboots, which forces a rebuild rather than clean-up.
- Stability risk: abusing kernel drivers can crash hosts, which can cause outages independent of encryption.
- Wider blast radius: EDR killers are often pushed by script or Group Policy to many hosts at once, so one compromised domain admin can blind a whole estate.
Detection
| Signal | Log source | Notes |
|---|---|---|
| Driver loaded (hash, signer, signature status) | Sysmon Event ID 6 | Match hashes and signers against LOLDrivers and the Microsoft blocklist; flag drivers loaded from user-writable or temp paths |
| New service with a kernel driver type | System log 7045; Security log 4697 | Rare on servers outside patch windows; correlate with the account that created it |
| Driver would have been blocked (audit mode) | Microsoft-Windows-CodeIntegrity/Operational 3076 | Use during blocklist pilots to find legitimate drivers that would break |
| Driver blocked (enforced) | Microsoft-Windows-CodeIntegrity/Operational 3077 | A block on a production host is a high-fidelity incident signal |
| Security agent stopped or heartbeat lost | EDR console, SIEM heartbeat rule | Alert on loss of heartbeat from a host that is still sending other logs (for example DNS or proxy) |
| Known EDR killer file or behaviour | EDR, antivirus PUA detection | ESET recommends detecting potentially unsafe applications and the killers themselves, not only the driver |
Correlation that works: new kernel service creation, followed within minutes by the EDR sensor stopping on the same host, followed by mass file changes or shadow copy deletion. Each event alone can be benign; the sequence is not. Because the agent may die first, forward driver load and service events to the SIEM in near real time through a separate channel such as Windows Event Forwarding. The EDR Guide covers sensor health monitoring in more detail.
Metrics
- Blocklist enforcement coverage: percentage of Windows endpoints and servers with the vulnerable driver blocklist enforced (not audit).
- Memory integrity coverage: percentage of capable devices reporting HVCI running.
- Blocklist currency: days since the deployed blocklist version matched Microsoft's latest downloadable list.
- Driver alert triage time: median time from a 3077 block or unknown kernel service to analyst review.
- Silent sensor gaps: count of hosts losing EDR heartbeat while other telemetry continues, per month.
Prevention and hardening
Controls compared
| Control | What it does | Limits |
|---|---|---|
| Microsoft vulnerable driver blocklist | Blocks drivers meeting Microsoft's criteria (kernel privilege escalation flaws, malicious drivers or certificates used for malware, behaviour circumventing the security model). On by default since the Windows 11 2022 update and enforced with HVCI, Smart App Control or S mode | In-box list updates less often than the downloadable list; running drivers only blocked after reboot; Microsoft states it does not guarantee to block every vulnerable driver |
| Memory integrity (HVCI) | Uses virtualisation-based security to enforce kernel code integrity | Incompatible drivers can cause failures, even boot failure; some vulnerable drivers still load |
| App Control for Business driver allowlist | Only approved drivers load at all | Effort to build and maintain; best for servers and Tier 0 |
| Removing local admin rights | Prevents most users from installing drivers | Does not stop an attacker holding domain or server admin |
| EDR tamper protection | Stops user-mode attempts to stop the agent | Kernel-level attacks can bypass it, which is why driver controls matter |
Rollout
- Inventory drivers. Export loaded driver lists from EDR or Sysmon Event ID 6 and compare with LOLDrivers and the Microsoft list to find what you actually run.
- Check memory integrity status. Microsoft documents the setting under Windows Security, Device security, Core isolation, Memory integrity, and reporting through the Win32_DeviceGuard WMI class.
- Pilot the downloadable blocklist in audit mode. Deploy the audit version of the policy to a pilot ring and collect CodeIntegrity 3076 events for two to four weeks. Microsoft documents copying the policy binary (SiPolicy.p7b) to %windir%\system32\CodeIntegrity and refreshing it.
- Fix or replace incompatible drivers. Update vendor drivers flagged in the pilot; do not create broad exceptions.
- Enforce in rings. Move to the enforced policy ring by ring, starting with servers that should have the fewest drivers. Reboot so running drivers are blocked.
- Enable memory integrity on capable devices after driver compatibility is confirmed.
- Add an App Control driver allowlist for Tier 0 (domain controllers, identity servers, backup servers). When merging Microsoft's block rules into your own policy, remove the Allow All rules as Microsoft instructs, or the merged policy allows everything.
- Schedule updates. Microsoft updates the list periodically; set a monthly check against the downloadable version.
Rollback: keep the previous policy version ready to redeploy. Microsoft documents a recovery path for memory integrity by setting the HypervisorEnforcedCodeIntegrity Enabled registry value to 0, and warns that a UEFI lock means Secure Boot must be disabled to recover. Test recovery on a pilot machine before you use the UEFI lock.
Decision guidance
- Small organisations: confirm memory integrity and the default blocklist are on, remove local admin from users, and rely on EDR tamper protection and heartbeat alerts.
- Mid-size organisations: add the downloadable blocklist with a quarterly or monthly update and SIEM rules for driver loads and kernel services.
- Large or high-risk organisations: full App Control driver allowlisting on Tier 0 and servers, plus hash and signer hunting with LOLDrivers data.
Where identity fits
Every BYOVD attack starts with an administrator identity. Local administrator passwords reused across hosts, domain admin accounts used for routine work and service accounts with local admin on every server are what let one compromise turn into estate-wide EDR killing. Unique local admin passwords (Windows LAPS), tiered admin accounts and removing interactive rights from service accounts all cut the number of hosts an attacker can load drivers on. See the Service Account Security Guide and the AD Domain Controller Hardening Guide.
Lessons from incidents
- Certificate theft becomes driver abuse. The NVIDIA certificate leak showed that expired signing certificates can still sign drivers Windows will load. Respond with certificate-based deny rules, not only hash blocks. See the Compromised Code Signing Certificate Playbook.
- Blocklist in audit mode forever. Pilots that never move to enforcement give a false sense of coverage.
- Exceptions for one old hardware tool. A single legacy utility driver allowed fleet-wide can be the exact driver an EDR killer brings.
- No alert on sensor death. Teams that only alert on detections, not on agent loss, see the attack only after encryption.
- Relying on HVCI alone. LOLDrivers data shows some vulnerable drivers still load with HVCI; layer the blocklist and allowlisting.
Framework mapping
| Control | NIST CSF 2.0 | NIST SP 800-53 Rev. 5 | MITRE D3FEND |
|---|---|---|---|
| Vulnerable driver blocklist | PR.PS-05 | CM-7 Least Functionality | D3-EDL Executable Denylisting |
| Driver allowlisting | PR.PS-05 | CM-7, CM-14 Signed Components | D3-EAL Executable Allowlisting |
| Memory integrity and code integrity | PR.PS-02 | SI-7 Software, Firmware, and Information Integrity | D3-DLIC Driver Load Integrity Checking |
| Driver load and sensor health monitoring | DE.CM-09 | SI-4 System Monitoring | D3-DLIC |
| Malware and PUA detection of EDR killers | DE.CM-09 | SI-3 Malicious Code Protection | D3-EDL |
Practitioner checklist
- Inventory loaded kernel drivers across endpoints and servers and compare them with the Microsoft blocklist and LOLDrivers.
- Confirm memory integrity is running on every capable device and report coverage monthly.
- Pilot the downloadable vulnerable driver blocklist in audit mode, review CodeIntegrity 3076 events, then enforce ring by ring.
- Add an App Control driver allowlist for domain controllers, identity servers and backup servers.
- Collect Sysmon Event ID 6, System 7045 and Security 4697 centrally and alert on unknown drivers and kernel services.
- Alert on EDR heartbeat loss when the host is still sending other telemetry.
- Remove local administrator rights from users and deploy Windows LAPS for local admin accounts.
- Block KEV-listed drivers (CVE-2015-2291, CVE-2018-19320, CVE-2018-19321, CVE-2021-21551) by hash and signer.
- Document a rollback path for each policy change and test HVCI recovery before using the UEFI lock.
Standards and references
- Microsoft Learn: Microsoft recommended driver block rules
- Microsoft Learn: Enable memory integrity
- MITRE ATT&CK: Exploitation for Privilege Escalation (T1068)
- MITRE D3FEND: Driver Load Integrity Checking
- CISA: #StopRansomware: Medusa Ransomware (AA25-071A)
- CISA: Known Exploited Vulnerabilities Catalog
- LOLDrivers: Living Off The Land Drivers
- ESET: Shifting sands, RansomHub's EDRKillShifter
Related NHI Mgmt Group resources: EDR Evasion and Defence Guide · Top 10 EDR Evasion Techniques · Application Allowlisting Guide · Ransomware Prevention Guide · DragonForce Ransomware Profile Guide