Skip to content

Platform Compatibility

CloudTaser's default agent requires an active BPF LSM and permission to run its node agent. A compiled CONFIG_BPF_LSM=y is insufficient if bpf is absent from /sys/kernel/security/lsm. Kernel capability, workload admission policy, and the agent's actual readiness are separate checks.

AKS requires a verified node configuration. Assessed stock Ubuntu and Azure Linux defaults omit bpf, so the default agent is expected to refuse (assessed, not run). Actual enforcement/refusal needs live evidence. The dated Azure Linux running-node record reports an active BPF LSM and conflicts with the vendor default. This page does not resolve that conflict or qualify a current AKS deployment.

Process-memory-access enforcement: BPF LSM matrix

The configuration audit is dated 2026-10-24, with a 2026-10-29 live COS test, and covers x86_64. It distinguishes live GKE measurements from vendor configuration inspection. A vendor config shows the kernel's compiled defaults; a boot-command-line override can change the active LSM list. It is not a live deployment qualification.

Provider / node OS Kernel examined BPF LSM evidence Kprobe override Secret memory Consequence
GKE Standard, COS 6.12.105+, measured; 6.6 / 6.12 / 6.18 vendor configs Active BPF LSM measured; vendor configs include bpf Disabled Enabled Tier 1 via lsm+fmod_ret, Ready with default strict settings on the measured node; qualify other images/releases
GKE Standard, Ubuntu 24.04.5 6.8.0-1062-gke, measured Absent from active LSM list, although built Enabled Enabled Refused with default requireLSM: true
GKE Autopilot COS COS config evidence above Disabled Enabled Unsupported for protected workloads: agent blocked by default; no CloudTaser allowlist qualification. Without a Ready agent the gate refuses workloads; no wrapper-only mode qualified
EKS Fargate Not qualified Required node agent cannot run Not available Not qualified Unsupported for protected workloads; default gate refuses workloads; no wrapper-only mode qualified
EKS AL2023 6.12.111-136.204 / 6.18.54-121.166 Official RPM configs include bpf in CONFIG_LSM Enabled Enabled Tier 1 + LSM capability; confirm active node stack and agent vectors
EKS Bottlerocket 6.12 / 6.18 kernel-kit configs Vendor configs include bpf in CONFIG_LSM Enabled Enabled Tier 1 + LSM capability; confirm active node stack and agent vectors
AKS Azure Linux 3 Vendor 3.0 kernel 6.6 config CONFIG_BPF_LSM=y, but default CONFIG_LSM excludes bpf Disabled Enabled Expected to refuse (assessed, not run) if the running LSM list omits bpf; verify the dated-record conflict on the actual node
AKS Ubuntu 24.04 / 22.04 6.8-azure / 5.15-azure in configuration assessment Assessment reports bpf absent; no live AKS qualification in this audit Not verified in this audit 6.8 built; 5.15 requires checking enablement Expected to refuse (assessed, not run) without active BPF LSM; actual enforcement/refusal needs live evidence
OpenShift / ROSA RHEL 9 5.14 / RHCOS 10 6.12 in configuration assessment Assessment reports bpf enabled; no live OpenShift qualification in this audit Assessment reports disabled 5.14 requires checking enablement Tier 2 candidate; verify exact node kernel, required-vector policy and privileged SCC

COS, AL2023, Bottlerocket and Azure Linux rows have the pinned vendor sources listed below. The AKS Ubuntu and OpenShift rows retain the configuration assessment's scope explicitly; they are not measured deployment guarantees. Self-managed k3s/RKE2, Talos, and other distributions need the same running-node checks. The container runtime name does not establish kernel support.

Measured GKE COS result

On 29 October 2026, the EU GKE Standard COS node (6.12.105+, N2D SEV-SNP) ran the following recorded release cohort and reported:

chart: 1.0.352
agent: v0.13.164
/status: tier=1, kernel_tier=1, enforcement_mode=lsm+fmod_ret
lsm_active=true, lsm_count=25, kprobes_active=false, override_available=false
/readyz: HTTP 200, ready=true, enforcement_profile=strict
accepted_gaps={}, unaccepted_gaps={}

The chart's default requireLSM=true and effective failClosedOnEnforcementGap=true were retained. fmod_ret supplies synchronous denial where this kernel lacks kprobe override; absence of kprobe override does not make this deployment Tier 2. This is an observed agent/readiness result for that cohort, not qualification of every COS image or every workload. Node-root tamper controls still reported incomplete coverage and audit-only strict BPF load policy; a Ready agent does not remove that boundary.

Measured GKE Ubuntu result

The hosted demo-lab node ran GKE 1.35.8-gke.1439001, UBUNTU_CONTAINERD (Ubuntu 24.04.5), kernel 6.8.0-1062-gke, containerd 2.2.7, and runc 1.2.8. The owner-authorized node inspection recorded:

/sys/kernel/security/lsm: lockdown,capability,landlock,yama,apparmor
CONFIG_BPF_LSM=y
CONFIG_LSM="landlock,lockdown,yama,integrity,apparmor"
CONFIG_BPF_KPROBE_OVERRIDE=y
CONFIG_SECRETMEM=y
lockdown: integrity

The boot command line had no lsm= override. The default agent refuses this node because BPF LSM is not active. Neither the Ubuntu label, confidential-node hardware, nor a compiled BPF LSM changes that result. A CI guest whose command line was changed to include bpf does not qualify the stock GKE image.

Synchronous-blocking eBPF security: per-distro matrix

Mode Required kernel capabilities What the mode establishes
Tier 1 + LSM Active BPF LSM plus effective synchronous paths through LSM, fmod_ret or kprobe override Synchronous paths for the configured vectors; verify actual attachments, self-tests and status
Tier 2 Active BPF LSM, with required secondary vectors lacking an effective synchronous path Covered procfs/ptrace paths remain protected; inspect the deployed agent's per-vector states and remaining detect-only paths
Kprobe-only, explicitly reduced Kprobe override without active BPF LSM Raw-path matching has symlink/bind-mount alias limitations; this is not the default supported posture

A detect-and-kill response can follow disclosure; it does not prove a read was prevented. The agent's current vector contract is authoritative for the deployed binary. In particular, a Tier 2 result does not promise readiness under a policy requiring every vector to deny synchronously. The default failClosedOnEnforcementGap policy can keep readiness false for required gaps. Do not silently relax that policy to make a platform table row pass.

requireLSM: false is not a general compatibility fix. Some shipped enforcing objects independently require the resolved BPF-LSM procfs gate. An explicit reduced-mode configuration must satisfy its exact agent contract and cannot be presented as equivalent protection.

Root resistance has a boundary

Node-root protection is tamper-evident and fail-closed, not absolute. Root with permitted module/BPF loading, kernel replacement, /proc/kcore access, or the ability to remove pins can act outside covered read APIs. Confidential computing addresses the external hypervisor, not arbitrary guest-kernel code. See Security Model.

Protection score by platform

A numerical score is not a deployment qualification. An image can have CONFIG_BPF_LSM=y and a high set of compiled capabilities while the active LSM stack still makes the default agent refuse startup. Confirm the runtime checks below before using the Protection Score Reference to compare additional controls.

Check your own nodes

Run these read-only checks through your approved node-access procedure:

cat /sys/kernel/security/lsm
uname -r
cat /proc/cmdline
grep -E 'CONFIG_(BPF_LSM|LSM|BPF_KPROBE_OVERRIDE|SECRETMEM)=' /boot/config-$(uname -r)

If the config is exposed only through /proc/config.gz, inspect that instead. A missing config file is missing evidence, not an enabled feature. Confirm bpf in the active LSM list, inspect the agent's startup/refusal diagnostics, and verify the required vector states and readiness on the exact released binary. Exec probes have a separate runtime-bootstrap limitation.

Managed node repair, scaling and upgrades can replace the image and invalidate local boot changes. This page does not recommend a privileged GRUB-edit/reboot workaround for managed nodes.

References