v0.4.0: offensive refocus — cloud pivots, CVE correlation, tier-0, escape chains by SnailSploit · Pull Request #5 · SnailSploit/KubeRoast

SnailSploit · 2026-06-01T03:06:32Z

Turns KubeRoast into an assessment tool an operator actually wants: concrete attacker primitives, multi-hop escalation chains, cloud-account pivots, and version-aware CVE correlation. Every finding carries the PoC command, blast radius, exploitability, and ATT&CK IDs. Compliance lint is hidden by default so attack paths lead.

Two commits: v0.3.0 (escape primitives, kubelet probes, workload coverage, manifest mode) and v0.4.0 (the offensive refocus below).

v0.4.0 — offensive depth

De-noise — "no kiddie stuff" by default

Defense-in-depth findings (POD-NO-LIMITS, POD-NO-APPARMOR, POD-NO-SECCOMP, POD-SATOKEN, SECRET-TLS-MANUAL, POLICY-NONE, PSS-NOT-ENFORCED) are hidden by default behind --include-hygiene. Default scans lead with escapes, chains, cloud pivots, and CVEs.

Cloud-account pivots (`scanners/cloud.py`)

The highest-impact real-world escalation — pod → cloud account — modeled as first-class findings:

CLOUD-IMDS-REACHABLE — per-namespace egress posture vs 169.254.169.254; auto-detects AWS/GCP/Azure from node providerID and emits the provider-specific IMDS credential-theft curl.
CLOUD-IRSA / CLOUD-GKE-WI / CLOUD-AZURE-WI — ServiceAccount→cloud-role federation from SA annotations, with the assume/exchange PoC.

CVE correlation (`scanners/cve.py` + `utils/versions.py`)

Offline, version-aware KB of high-impact escape/escalation CVEs keyed on apiserver / kubelet / runtime / kernel / ingress-nginx versions. CISA-KEV + EPSS tags verified against live feeds:

CVE	EPSS	KEV	Component
CVE-2025-1974 (IngressNightmare)	0.91	—	ingress-nginx < 1.11.5 / 1.12.1
CVE-2018-1002105 (apiserver proxy)	0.90	—	k8s < 1.10.11/1.11.5/1.12.3
CVE-2024-1086 (nf_tables LPE)	0.85	✅	kernel < 6.8
CVE-2024-21626 (Leaky Vessels)	0.05	—	runc ≤ 1.1.11
CVE-2019-5736 (runc overwrite)	0.59	—	runc < 1.0
CVE-2022-0185 (fsconfig)	—	✅	kernel < 5.16.2
CVE-2020-15257, CVE-2021-25741, CVE-2020-8559, CVE-2023-5044, CVE-2021-25742	—	—	containerd / kubelet / apiserver / ingress-nginx

Tier-0 / crown-jewel identification (`attackpaths/tier_zero.py`)

AP-TIER0 names principals that are effectively cluster-admin without the literal binding: CSR create+approve (mint a system:masters cert), nodes/proxy, cluster-wide pods/exec, read-all-secrets, serviceaccounts/token, webhook config writes, impersonate/escalate/bind, nodes patch. Skips expected holders (kube-system, system:masters).

CSR cert-minting in multi-hop chains

sa_chains adds the create CSR + approve → system:masters edge, so escalation chains can terminate via certificate minting.

Hardcoded creds in pod specs (`scanners/podsecrets.py`)

ENV-LEAK-* / ARG-LEAK-* — literal credentials in container env values and command/args. secretKeyRef references are never flagged; values are redacted in output.

CLI / plumbing

--skip-cloud, --skip-cve, --include-hygiene. Nodes listed once and reused for probes + cloud + CVE. New listers (serviceaccounts, networkpolicies) + server-version getter; Admissionregistration + Batch API clients.

Demo (manifest mode)

A synthetic cluster with an IRSA SA holding create+approve CSR, an ingress-nginx v1.11.0 deployment, and a pod with hardcoded env creds produces — critical findings first:

[CRITICAL] trivial  Escalation chain (1 hop): sa:prod:ci-deployer → cluster-admin   (CSR cert-mint)
[CRITICAL] trivial  Effective cluster-admin (tier-0): sa:prod:ci-deployer
[CRITICAL] easy     ingress-nginx 'IngressNightmare' RCE   [EPSS=0.91]
[MEDIUM]   trivial  Likely hardcoded credential in pod env: DATABASE_PASSWORD
...                 CLOUD-IRSA: SA federated to arn:aws:iam::...:role/ci-deployer-role

…and POD-NO-LIMITS / POD-NO-SECCOMP / POD-NO-APPARMOR / POD-SATOKEN are correctly suppressed.

Test plan

pytest tests/ — 200/200 passing (was 38 at the start of this PR)
New suites: test_cloud (16), test_cve (24), test_tier_zero (11), test_podsecrets (10), test_cli (6 — hygiene gating + exit codes)
End-to-end manifest demo verifies offensive findings lead and hygiene is hidden by default; --include-hygiene brings it back
--fail-on, --min-exploitability, --version, --no-color exercised
Manual live-cluster run (no cluster harness in this repo)

https://claude.ai/code/session_01Eq8TDiVnuuVQU2HR9KtCNg

Rewrites the tool from compliance checklist into operator-grade red team recon. Every finding now carries the PoC command an attacker would run, the concrete blast radius (impact), exploitability score, and ATT&CK IDs. New offensive scanners ---------------------- - escapes.py Container escape primitive analysis. Per-pod scoring with the exact exploit command for privileged+hostPID, docker.sock, containerd.sock, /var/lib/etcd, /etc/kubernetes, SYS_ADMIN, SYS_MODULE, SYS_PTRACE (rated easy/hard depending on hostPID), SYS_RAWIO, DAC_READ_SEARCH, NET_ADMIN, writable /proc/sys/kernel/core_pattern, hostNetwork, hostIPC, and the root-no-drop multiplier. - kubelet.py Active HTTP probes against /pods, /metrics, /configz on 10250 and 10255. Distinguishes anonymous-allow (critical), anonymous-but-authz (medium), proper 401 (info — good). - tokens.py JWT decode of SA token Secrets. Flags legacy non-expiring tokens as persistence risk; surfaces non-default audiences (IRSA / Workload Identity). - configmaps.py Credential hunt in ConfigMaps — high-signal regexes for AWS AKIDs, PEM keys, GCP service-account JSON, GitHub PATs, Slack tokens, JDBC URLs, Azure connection strings. Redacts long matches. - webhooks.py Admission webhook bypass audit — failurePolicy=Ignore, long timeouts, sideEffects=Unknown, missing caBundle. New multi-hop attack path module -------------------------------- - attackpaths/sa_chains.py BFS over an escalation graph with edges for create-pod-as-SA, exec-into-SA-pod, patch-deployment-as-SA, create/update RoleBindings, escalate verb, impersonate, and serviceaccounts/token. Emits one finding per chain with per-hop PoC commands. Workload coverage ----------------- - workloads.py Extracts pod templates from Deployment, StatefulSet, DaemonSet, Job, CronJob, ReplicaSet, ReplicationController. Findings get relabeled from pod/X to <kind>/X for clear attribution. Shift-left manifest mode ------------------------ - manifests.py YAML/JSON loader with camelCase→snake_case (including k8s python client quirks like external_i_ps). --manifests <dir> runs the same rules without API access. Finding model ------------- - New fields: exploitability, exploit, impact, attack_techniques. CLI --- - ASCII banner, ANSI colors with NO_COLOR support, --no-color flag, --version, --min-exploitability filter, --skip-kubelet-probe, --skip-configmaps, --skip-webhooks, richer --help with examples. - Text reporter shows severity histogram, exploitability breakdown, and per-finding PoC blocks. Tests ----- - 38 → 134 (3.5×). New suites: test_escapes (24), test_workloads (8), test_configmaps (17), test_tokens (7), test_webhooks (9), test_sa_chains (8), test_manifests (11), test_kubelet (12) with a real local HTTP server so HTTP code paths actually run. Docs ---- - README rewritten around the operator playbook: sample finding output, 5-minute triage workflow, full scanner table, comparison vs kube-bench / kubescape / kube-hunter, snailsploit.com docs link prominent. https://claude.ai/code/session_01Eq8TDiVnuuVQU2HR9KtCNg

Sharpens the tool toward real assessment tradecraft and cuts the compliance lint that was burying signal. De-noise ("no kiddie stuff") ---------------------------- Defense-in-depth findings (POD-NO-LIMITS, POD-NO-APPARMOR, POD-NO-SECCOMP, POD-SATOKEN, SECRET-TLS-MANUAL, POLICY-NONE, PSS-NOT-ENFORCED) are now hidden by default behind --include-hygiene. Default scans lead with escape primitives, escalation chains, cloud pivots, and CVEs. Cloud-account pivots (scanners/cloud.py) ---------------------------------------- - CLOUD-IMDS-REACHABLE: per-namespace egress posture vs 169.254.169.254. Auto-detects AWS/GCP/Azure from node providerID and emits the exact provider-specific IMDS credential-theft curl. - CLOUD-IRSA / CLOUD-GKE-WI / CLOUD-AZURE-WI: ServiceAccount→cloud-role federation from SA annotations, with the assume/exchange PoC. Models the pod -> cloud account pivot as a first-class finding. CVE correlation (scanners/cve.py + utils/versions.py) ----------------------------------------------------- Offline, version-aware knowledge base of high-impact escape/escalation CVEs keyed on apiserver / kubelet / runtime / kernel / ingress-nginx versions. Carries CISA-KEV status and EPSS (verified against live feeds). Includes IngressNightmare (CVE-2025-1974, EPSS 0.91), the apiserver-proxy bug (CVE-2018-1002105), runc Leaky Vessels (CVE-2024-21626), nf_tables LPE (CVE-2024-1086, KEV), fsconfig (CVE-2022-0185, KEV), containerd-shim, subPath, and the ingress-nginx annotation-injection chain. Tier-0 / crown-jewel identification (attackpaths/tier_zero.py) -------------------------------------------------------------- AP-TIER0 names principals that are effectively cluster-admin without the literal binding: CSR create+approve (mint a system:masters cert), nodes/proxy, cluster-wide pods/exec, read-all-secrets, serviceaccounts/token, webhook config writes, impersonate/escalate/bind, nodes patch. Skips expected holders (kube-system, system:masters). Multi-hop chains now route through CSR cert-minting --------------------------------------------------- sa_chains adds the CSR create+approve -> system:masters edge, so escalation chains can terminate via certificate minting. Hardcoded credentials in pod specs (scanners/podsecrets.py) ----------------------------------------------------------- ENV-LEAK-* / ARG-LEAK-*: literal credentials in container env values and command/args, reusing the high-signal credential patterns. secretKeyRef references are never flagged. Values are redacted in output. CLI / plumbing -------------- - New flags: --skip-cloud, --skip-cve, --include-hygiene. - Nodes listed once and reused for probes + cloud + CVE. - kube.py: serviceaccount, networkpolicy listers + server-version getter; AdmissionregistrationV1Api + BatchV1Api clients. Tests ----- 134 -> 200. New suites: test_cloud (16), test_cve (24), test_tier_zero (11), test_podsecrets (10), test_cli (6, hygiene gating + exit codes). Docs ---- README reframed around offensive coverage: cloud-pivot and CVE tables, tier-0, the no-kiddie-stuff default, updated comparison and RBAC needs. https://claude.ai/code/session_01Eq8TDiVnuuVQU2HR9KtCNg

Control-plane probes (scanners/controlplane.py) ----------------------------------------------- Active, read-only probes of control-plane nodes — the crown jewels that bypass Kubernetes RBAC entirely: - CP-ETCD-NOAUTH (critical): etcd 2379 answers /version unauthenticated → client-cert auth is off → read/write every Secret directly via etcdctl, invisible to API audit logs. Emits the exact etcdctl dump command. - CP-ETCD-REACHABLE (high): etcd client port reachable from the scan origin even when mTLS is enforced — it should never be routable from the workload network. - CP-ETCD-PEER (medium): etcd peer port 2380 reachable. - CP-APISERVER-INSECURE (critical): legacy insecure apiserver port 8080 serves the API unauthenticated → implicit cluster-admin. - CP-PORT-8080-OPEN (low): 8080 open but unconfirmed — identify listener. Probing prefers control-plane-labeled nodes (etcd is co-located on kubeadm/self-managed); on managed control planes (EKS/GKE/AKS) the ports aren't reachable and nothing fires. Reuses the kubelet TCP/HTTP probe helpers; gated by --skip-kubelet-probe (all active HTTP probing). DaemonSet escape amplification (utils/workloads.py) --------------------------------------------------- A DaemonSet runs a pod on every node, so an escape primitive in its template is a foothold on ALL nodes, not one pod. Escape/AttackPath findings sourced from a DaemonSet template now get their impact re-framed to node-wide blast radius. Applied in both cluster and manifest mode. Tests ----- 200 -> 215. New test_controlplane (13) covering etcd noauth/reachable/peer, plain-HTTP etcd, insecure apiserver, target selection, and end-to-end; plus DaemonSet amplification tests in test_workloads. Docs ---- README: new Control-Plane Exposure coverage section, ControlPlane category, DaemonSet-amplification note, updated comparison + architecture. https://claude.ai/code/session_01Eq8TDiVnuuVQU2HR9KtCNg

claude added 2 commits June 1, 2026 03:05

SnailSploit changed the title ~~v0.3.0: offensive-grade scanners, multi-hop escape chains, killer CLI~~ v0.4.0: offensive refocus — cloud pivots, CVE correlation, tier-0, escape chains Jun 1, 2026

This was referenced Jun 1, 2026

NetworkPolicy lateral-movement gap analysis (flat-network namespaces) #6

Open

v0.3.0: industry-standard scanner — CIS/MITRE/CWE, SARIF, offline manifests, CI #3

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

v0.4.0: offensive refocus — cloud pivots, CVE correlation, tier-0, escape chains#5

v0.4.0: offensive refocus — cloud pivots, CVE correlation, tier-0, escape chains#5
SnailSploit wants to merge 3 commits into
mainfrom
claude/inspiring-rubin-l3BeQ

SnailSploit commented Jun 1, 2026 •

edited

Loading

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

SnailSploit commented Jun 1, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

v0.4.0 — offensive depth

De-noise — "no kiddie stuff" by default

Cloud-account pivots (scanners/cloud.py)

CVE correlation (scanners/cve.py + utils/versions.py)

Tier-0 / crown-jewel identification (attackpaths/tier_zero.py)

CSR cert-minting in multi-hop chains

Hardcoded creds in pod specs (scanners/podsecrets.py)

CLI / plumbing

Demo (manifest mode)

Test plan

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

SnailSploit commented Jun 1, 2026 •

edited

Loading

Cloud-account pivots (`scanners/cloud.py`)

CVE correlation (`scanners/cve.py` + `utils/versions.py`)

Tier-0 / crown-jewel identification (`attackpaths/tier_zero.py`)

Hardcoded creds in pod specs (`scanners/podsecrets.py`)