Root Attack Surface

Authorised application boundary

Authorised application code is trusted. An authorised app can send out a secret it has copied or encoded. Current content checks do not provide a general guarantee against this disclosure or a malicious authorised app. Memory-read tests and a Ready agent do not prove copied/encoded-secret exfiltration prevention. Enforced NetworkPolicy can restrict destinations, but allowed traffic remains a channel. See Security Model.

This page analyzes the attack surface when the adversary has root access on the Kubernetes node. This is the primary threat model for cloudtaser: the cloud provider has root-equivalent access through hypervisor management, support tooling, and legal compulsion under the CLOUD Act.

Executive Summary

On Linux 5.14+ with BPF LSM enforcement active, cloudtaser blocks every known software-level attack path available to a root user on the Kubernetes node -- including the cloud provider operating under CLOUD Act compulsion. memfd_secret removes secret pages from the kernel direct map, and the BPF LSM hook lsm_ptrace_access_check synchronously denies process_vm_readv, ptrace, and /proc access against monitored processes. Enforcement requires bpf in the kernel's boot-time LSM stack; this has been verified on GKE Confidential Computing / COS (COS 6.12, cloudtaser-ebpf v0.4.54). Other managed Kubernetes platforms may support BPF LSM but have not been independently confirmed by CloudTaser -- verify per platform. The only remaining exposure is the hypervisor's physical memory view, which requires confidential computing hardware to close. On kernels without BPF LSM, these vectors are not blocked -- the read succeeds and the secret is disclosed before any reactive kill fires. See Platform Compatibility for the full matrix.

---

Threat Model

The Adversary

The adversary has full root access on the Kubernetes worker node. This includes:

• Cloud provider operations staff -- via hypervisor console, management plane APIs, support SSH access
• Cloud provider under legal compulsion -- CLOUD Act warrant, FISA 702 directive, National Security Letter
• An attacker who escalated to root -- through container escape, kernel exploit, or compromised node agent

The Goal

Extract secrets from a running cloudtaser-protected process. The secrets are:

• Database credentials, API keys, encryption keys
• Fetched from an EU-hosted OpenBao instance
• Stored exclusively in process memory (never on disk, never in etcd)

The Question

What can a root attacker actually do, and what does cloudtaser block?

---

Defense Layer 1: Wrapper (Process-Level)

The wrapper applies six memory protection mechanisms before launching the application:

Defense	What It Does	Can Root Bypass?	Kernel Requirement
memfd_secret	Removes pages from kernel direct map	No -- hardware-level guarantee	Linux 5.14+
mlock	Pins pages in RAM, prevents swap	No (but root can read RAM via other paths)	Any Linux
MADV_DONTDUMP	Excludes from core dumps	Yes -- can write to coredump_filter	Linux 3.4+
PR_SET_DUMPABLE(0)	Blocks ptrace, restricts /proc	Yes -- can load kernel module to override	Any Linux
Wrapper environ clean (environ_scrubbed)	Wrapper exec'd without secrets in envp[]; /proc/self/environ snapshot frozen at exec time before vault fetch	Structural -- no secret bytes ever written to wrapper environ region	Any Linux
LD_PRELOAD interposer	Returns memfd pointers from getenv() for glibc-dynamic apps; env-var inheritance still works for static binaries	N/A -- prevents heap copies on the dynamic-link path	Any Linux

memfd_secret is the critical defense

memfd_secret removes a direct-map alias. It is a kernel primitive, not hardware isolation against arbitrary kernel code. Ordinary application copies and raw direct store login remain outside its boundary.

---

Defense Layer 2: eBPF Agent (Enforcement)

Even with memfd_secret, a root attacker may attempt secondary attacks. The eBPF agent blocks these in real time.

Secret Exfiltration Attempts

Attack	Syscall	Enforcement	Blocked?
Write secret to file	write, writev	kprobe block (content matching)	Yes
Send secret over network	sendto, sendmsg	kprobe block (content matching)	Yes
Zero-copy exfiltration	sendfile, splice, tee, vmsplice	kprobe block	Yes
Async I/O bypass	io_uring_setup	kprobe block	Yes

Process Memory Access Attempts

Attack	Syscall / Path	Enforcement	Blocked?
Read /proc/PID/mem	openat	kprobe block	Yes
Read /proc/PID/environ	openat	kprobe block (all monitored PIDs in cgroup-array, including child PIDs in sibling cgroups)	Yes
Cross-process read	process_vm_readv	BPF LSM hook (lsm_ptrace_access_check)	Blocked where BPF LSM active (verified on GKE CC/COS); not blocked where kernel lacks bpf in LSM stack -- read succeeds, reactive kill fires after disclosure
Cross-process write	process_vm_writev	BPF LSM hook (lsm_ptrace_access_check)	Blocked where BPF LSM active (verified on GKE CC/COS); not blocked otherwise -- write succeeds, reactive kill fires after disclosure
Debug attach	ptrace ATTACH/SEIZE	BPF LSM hook (lsm_ptrace_access_check)	Blocked where BPF LSM active (verified on GKE CC/COS); not blocked otherwise -- attach succeeds, reactive kill fires after disclosure
Performance sampling	perf_event_open	kprobe block	Yes
Page fault interception	userfaultfd	kprobe block	Yes
Raw memory device	/dev/mem, /dev/kmem, /proc/kcore	kprobe block	Yes

Information Disclosure Attempts

Attack	Path	Enforcement	Blocked?
Memory layout	/proc/PID/maps, pagemap, smaps	kprobe block	Yes
Register values	/proc/PID/syscall	kprobe block	Yes
Kernel stack trace	/proc/PID/stack	kprobe block	Yes
Process metadata	/proc/PID/cmdline, status, wchan	kprobe block	Yes
Re-enable core dumps	/proc/PID/coredump_filter write	kprobe block	Yes

Privilege Escalation Attempts

Attack	Syscall	Enforcement	Blocked?
Load kernel module	init_module, finit_module	Global detection + monitored PID block	Detected / Blocked
Load eBPF program	bpf(BPF_PROG_LOAD)	Global detection + monitored PID block	Detected / Blocked

---

Defense Layer 3: Reactive Kill

Reactive kill is not a substitute for a synchronous deny path: the default chart does not count a vector covered only by reactive kill as enforced, and the agent reports HTTP 503 while any required vector is unenforced. See Kernel Compatibility.

On kernels where BPF LSM is not active in the boot-time LSM stack (i.e., bpf is absent from lsm=), the eBPF agent falls back to reactive enforcement for process-memory-access vectors (process_vm_readv, ptrace, /proc/PID/mem):

1. The syscall executes (tracepoint fires on exit)
2. The eBPF program detects the violation
3. SIGKILL is sent to the offending process immediately

Reactive kill does NOT prevent secret disclosure

On kernels without BPF LSM, the process_vm_readv / ptrace / /proc/PID/mem call completes and returns the secret before the reactive kill fires. The attacker already has the data. CloudTaser detects the access and terminates the offending process after the fact, but this does not prevent the extraction. For synchronous protection, use a kernel with bpf in the boot-time LSM stack (verified on GKE Confidential Computing / COS). See Platform Compatibility.

---

Known Limitations

Root Loading a Kernel Module (Pre-5.14 Kernels)

The fundamental limitation on pre-5.14 kernels

A root attacker can load a custom kernel module from a non-monitored PID. The module runs with full kernel privileges and can scan all process memory.

On pre-5.14 kernels:

• cloudtaser detects the module load via global privilege escalation detection
• cloudtaser logs the event and can alert operators
• cloudtaser cannot block module loads from non-monitored PIDs without breaking system services (kubelet, kube-proxy, container runtime all load modules)
• The loaded module can scan process memory because memfd_secret is unavailable

On 5.14+ kernels:

• The same module load is detected
• The missing direct-map alias limits ordinary access; arbitrary kernel code is outside the isolation guarantee
• The attack fails at the hardware level

Recommendation: Always target Linux 5.14+ for production deployments.

Hibernate / Suspend-to-Disk

Full RAM contents are written to disk during hibernate. This is a kernel-level operation that:

• Bypasses mlock (pages are force-written to the hibernate image)
• Bypasses eBPF (no syscall to intercept)
• Bypasses memfd_secret (the kernel reads physical pages directly for the hibernate image)

Disable hibernate on cloudtaser nodes

# Verify hibernate is disabled
cat /sys/power/state
# Should show: freeze mem
# Should NOT show: disk

io_uring for Application Performance

cloudtaser blocks io_uring_setup() for monitored processes because io_uring's submission queue bypasses per-syscall eBPF inspection. Applications that depend on io_uring for performance (high-throughput databases, network proxies) must use standard syscalls.

This is an intentional trade-off: security over performance for secret-handling workloads.

Confidential Computing (Not Yet Integrated)

The hypervisor has physical access to VM memory. On standard VMs, this means:

• The cloud provider can inspect memory through the hypervisor
• DMA attacks from host devices can read VM memory
• Live migration exposes memory contents

These attacks require confidential computing hardware (AMD SEV-SNP, Intel TDX, ARM CCA) to prevent. cloudtaser's software protections are necessary but not sufficient against a hypervisor-level adversary. Confidential computing integration is on the roadmap.

---

Attack Tree Summary

The following table shows what a root attacker can achieve at each kernel level:

Attack Path	Pre-5.14, No eBPF	Pre-5.14, With eBPF	5.14+, With eBPF	5.14+, eBPF + CC
/proc/PID/mem	Read secrets	Blocked	Blocked + memfd	Blocked + memfd + encrypted VM
/proc/PID/environ (child)	Read secrets	Blocked by eBPF kprobe	Blocked + memfd	Blocked + memfd + encrypted VM
ptrace	Read secrets	Blocked	Blocked + memfd	Blocked + memfd + encrypted VM
process_vm_readv	Read secrets	Blocked (BPF LSM) or read succeeds (no LSM; reactive kill after disclosure)	Blocked (BPF LSM) + memfd	Blocked (BPF LSM) + memfd + encrypted VM
Kernel module scan	Read secrets	Detected (not blocked)	Detected, but memfd immune	Fully blocked
/dev/mem	Read secrets	Blocked	Blocked + memfd	Blocked + memfd + encrypted VM
Core dump	Read secrets	MADV_DONTDUMP	MADV_DONTDUMP + memfd	MADV_DONTDUMP + memfd + encrypted VM
Swap forensics	mlock protects	mlock protects	mlock protects	mlock + encrypted VM
Hypervisor inspect	Read secrets	Read secrets	Read secrets	Encrypted

Read the table right to left

Each column adds a defense layer. The columns describe particular read paths. They do not cover raw direct store login, arbitrary guest-kernel code or authorised application disclosure. Confidential compute addresses the external hypervisor separately; no column establishes that all attacks are blocked.

---

Protection Guarantees

Enforcement depends on kernel capabilities

The guarantees below assume BPF LSM is active in the kernel's boot-time LSM stack (lsm=...,bpf). This has been verified on GKE Confidential Computing / COS (COS 6.12 with cloudtaser-ebpf v0.4.54). Other managed Kubernetes platforms (EKS, AKS) may support BPF LSM but have not been independently confirmed by CloudTaser -- verify on your platform before relying on synchronous blocking. On kernels without BPF LSM, process-memory-access vectors (process_vm_readv, ptrace, /proc/PID/mem) are not blocked -- the read succeeds and the secret is disclosed. CloudTaser detects the access and terminates the offending process after the fact, but this does not prevent extraction. See Platform Compatibility for the full enforcement matrix.

On Linux 5.14+ with BPF LSM active (recommended):

• Active BPF LSM denies covered procfs/ptrace read paths; verify each deployed vector separately
• Arbitrary guest-kernel code, ordinary application copies and raw direct store login remain limits
• External hypervisor resistance additionally requires qualified confidential compute

On Linux 5.14+ without BPF LSM (detect-only mode -- NOT protected against process-memory reads):

• memfd_secret still protects secret pages at the hardware level -- they are absent from the kernel direct map regardless of enforcement mode
• process_vm_readv, ptrace, and /proc/PID/mem access attempts are not blocked -- the read completes and the secret is returned to the attacker before any reactive action fires
• CloudTaser detects the access via tracepoint and terminates the offending process with SIGKILL, but this is post-hoc -- it does not prevent the initial disclosure
• For synchronous protection of process-memory-access vectors, use a kernel with BPF LSM active (e.g., GKE Confidential Computing / COS)

On pre-5.14 kernels with eBPF enforcement:

• Covered vectors require individual enforcement evidence; authorised copied/encoded-secret disclosure remains outside the guarantee
• A determined root attacker who loads a custom kernel module can scan memory
• This attack is detected and logged but not prevented
• mlock, MADV_DONTDUMP, and environment scrubbing remain effective

Without eBPF enforcement:

• Memory protections (memfd_secret, mlock, MADV_DONTDUMP, PR_SET_DUMPABLE) are still active
• No runtime monitoring of attack attempts
• Root can use /proc, ptrace, and process_vm_readv (blocked only by PR_SET_DUMPABLE, which root can override)
• Not recommended for production

Security Model | Protection Score