LID

两个内核子系统。不兼容的信任假设。一个漏洞。

<br>

---

## LID-001: eBPF 路径名重写

**重点发现。** 一个在 `do_sys_openat2` 上的 BPF kprobe 在用户内存中重写了文件名，然后再由内核复制。AppArmor 检查了重写后的路径，授予了访问权限。零审计追踪。```
  process: open("/tmp/secret.txt")
       │
       ▼
  do_sys_openat2()
       │
    ★ LID kprobe fires here
    │  bpf_probe_write_user()
    │  rewrites "/tmp/secret.txt" → "/tmp/.bypass_link"
       │
       ▼
  getname_flags()          ← kernel copies the (rewritten) path
       │
       ▼
  security_file_open()     ← LSM hooks check "/tmp/.bypass_link"
       │                      AppArmor: ALLOW ✓
       ▼
  VFS opens inode          ← same file content (hard link)
       │
       ▼
  return fd to process     ← success, zero audit trace

### 演示输出```
╔══════════════════════════════════════════════════════════╗
║  Phase 1: AppArmor ENFORCING — access should be DENIED   ║
╚══════════════════════════════════════════════════════════╝

  $ /tmp/test_reader
  [-] DENIED: open() failed: Permission denied (errno=13)

╔══════════════════════════════════════════════════════════╗
║  Phase 2: Loading LID — BPF kprobe pathname rewrite       ║
╚══════════════════════════════════════════════════════════╝

  [*] BPF kprobe attached to do_sys_openat2

╔══════════════════════════════════════════════════════════╗
║  Phase 3: With LID active — access should be GRANTED      ║
╚══════════════════════════════════════════════════════════╝

  $ /tmp/test_reader
  [+] SUCCESS: Read 44 bytes: SECRET_DATA=this_is_protected_content_12345

╔══════════════════════════════════════════════════════════╗
║  Phase 4: Stealth check — audit log inspection             ║
╚══════════════════════════════════════════════════════════╝

  $ dmesg | grep apparmor | grep DENIED
  (empty — no denial was ever generated)

**漏洞位置：** `io_uring/msg_ring.c`, `io_msg_install_complete()` — 直接调用 `__io_fixed_fd_install()`，跳过了 LSM 钩子。

**通过 ftrace 验证：** `security_file_receive` 对 SCM_RIGHTS 触发，但对 MSG_RING 不触发。

**影响范围：** Linux 5.18+ 至 v7.1-rc3（截至 2026-05-17 未修复）。

**PoC:** [`findings/lid-002-iouring-msgring/msg_ring_bypass.c`](https://github.com/azqzazq1/lid/blob/HEAD/findings/lid-002-iouring-msgring/)

<br>

---

## LID-003：新挂载 API 绕过 AppArmor

新的挂载 API（`fsopen` + `fsconfig` + `fsmount` + `move_mount`）**从未调用 `security_sb_mount()`**——这是 AppArmor 实现的唯一挂载钩子。```
  OLD API:  mount("proc", "/mnt", "proc", 0, NULL)
            → security_sb_mount()  → AppArmor: DENY ✗

  NEW API:  fsopen("proc") → fsconfig(CMD_CREATE) → fsmount() → move_mount()
            → security_sb_kern_mount()  → AppArmor: (not implemented) → ALLOW ✓

在Ubuntu/Debian上，一个具有bpffs委托的容器可以：
- 创建BPF令牌 → AppArmor看不到任何东西
- 使用令牌加载跟踪程序 → AppArmor看不到任何东西
- 委托CAP_PERFMON → 禁用Spectre缓解 → AppArmor看不到任何东西

**详情：** [`findings/lid-004-bpf-token/`](https://github.com/azqzazq1/lid/blob/HEAD/findings/lid-004-bpf-token/)

<br>

---

## LID-005: 来自默认Docker容器的AF_XDP tc出口绕过

AF_XDP的拷贝模式传输路径（`xsk_generic_xmit` → `__dev_direct_xmit`）**绕过容器接口上的tc出口分类器**。已使用tc u32 DROP-ALL验证——AF_PACKET被阻止，AF_XDP通过。

**然而：** 使用Cilium v1.19（kind集群，拒绝所有出口NetworkPolicy）进行测试——**Cilium未被绕过。** Cilium在节点侧veth对端入口（通过tcx/ingress的`cil_from_container`）实施强制，并具有源IP验证。AF_XDP数据包在`bpf_lxc.c:1603`处作为"Invalid source ip"被丢弃。影响仅限于仅依赖tc出口分类器进行网络策略的环境（纯Docker + tc规则，无CNI）。```
  Normal TX path:
  socket → dev_queue_xmit() → sch_handle_egress() → tc classifiers → driver
                                     ↑
                                  ENFORCED (Cilium eBPF, Calico, tc u32)

  AF_XDP copy-mode TX:
  xsk_sendmsg() → xsk_generic_xmit() → __dev_direct_xmit() → driver
                                              ↑
                                           SKIPPED (tc never runs)

数据包携带攻击者控制的以太网头部——伪造的MAC地址、伪造的IP地址——尽管有活动的DROP策略，仍能到达Docker桥接。

**PoC + 详情：** [`findings/lid-005-afxdp-tc-bypass/`](https://github.com/azqzazq1/lid/blob/HEAD/findings/lid-005-afxdp-tc-bypass/)

<br>

---

## 架构：为什么BPF LSM无法修复此问题```
  LSM Hook Chain: call_int_hook(file_open, ...)

  ┌─────────────┐   ┌─────────────┐   ┌─────────────┐   ┌─────────────┐
  │  Lockdown    │──>│  Capability  │──>│   AppArmor   │──>│  BPF LSM    │
  │  return 0    │   │  return 0    │   │  return -13  │   │  (never     │
  │  (allow)     │   │  (allow)     │   │  (DENY) ██   │   │   reached)  │
  └─────────────┘   └─────────────┘   └──────┬───────┘   └─────────────┘
                                              │
                                         loop breaks
                                         RC = -EACCES

  ★ The LSM framework is correct. No module can undo another's denial.
  ★ LID operates OUTSIDE the framework — before hooks run, or where hooks don't exist.

               ┌───────────────────────────────────────┐
               │          THE ATTACK TIMELINE           │
               │                                       │

<br>

## 项目结构```
LID/
├── src/
│   ├── bpf/
│   │   └── lid.bpf.c              # BPF kprobe — pathname rewriter (LID-001)
│   └── loader/
│       └── lid_loader.c           # Userspace loader + event monitor
├── findings/
│   ├── lid-001-ebpf-pathname/     # eBPF pathname rewriting bypass
│   ├── lid-002-iouring-msgring/   # io_uring MSG_RING missing LSM hook
│   │   ├── msg_ring_bypass.c      # PoC with ftrace verification
│   │   └── ADVISORY.md            # Technical advisory
│   ├── lid-003-mount-api/         # New mount API AppArmor bypass
│   ├── lid-004-bpf-token/        # BPF token AppArmor/Smack zero coverage
│   └── lid-005-afxdp-tc-bypass/  # AF_XDP tc egress bypass from container
├── tests/
│   └── test_reader.c              # Victim binary for LID-001 demo
├── scripts/
│   ├── check_prerequisites.sh     # Verify system requirements
│   ├── setup_env.sh               # Install build dependencies
│   ├── setup_demo.sh              # Create demo environment
│   ├── run_demo.sh                # Run full LID-001 demonstration
│   └── teardown.sh                # Clean up everything
├── docs/
│   └── RESEARCH.md                # Full technical research paper
├── publish/                       # Articles for Medium, dev.to, etc.
├── Makefile
├── LICENSE
├── SECURITY.md
└── README.md