diff --git a/stacks/crowdsec/main.tf b/stacks/crowdsec/main.tf
index a4c0013f..ca768356 100644
--- a/stacks/crowdsec/main.tf
+++ b/stacks/crowdsec/main.tf
@@ -32,4 +32,8 @@ module "crowdsec" {
   # Same key the traefik-stack bouncer middleware uses — seeded into LAPI so the
   # bouncer authenticates and pulls decisions (was unregistered → 403 → fail-open).
   ingress_bouncer_key = data.vault_kv_secret_v2.secrets.data["ingress_crowdsec_api_key"]
+  # Real enforcement replacing the dead Traefik plugin: kvsync feeds the proxied
+  # edge Worker via Cloudflare KV; firewall is the direct-host nftables bouncer.
+  kvsync_bouncer_key   = data.vault_kv_secret_v2.secrets.data["kvsync_bouncer_key"]
+  firewall_bouncer_key = data.vault_kv_secret_v2.secrets.data["firewall_bouncer_key"]
 }
diff --git a/stacks/crowdsec/modules/crowdsec/firewall_bouncer.tf b/stacks/crowdsec/modules/crowdsec/firewall_bouncer.tf
new file mode 100644
index 00000000..c161486d
--- /dev/null
+++ b/stacks/crowdsec/modules/crowdsec/firewall_bouncer.tf
@@ -0,0 +1,254 @@
+# =============================================================================
+# cs-firewall-bouncer — in-kernel (nftables) enforcement DaemonSet
+# =============================================================================
+# CrowdSec currently enforces NOTHING at the network layer: the Traefik Yaegi
+# (lua) bouncer plugin is dead, so banned IPs still reach Traefik. For DIRECT
+# (non-Cloudflare-proxied) hosts we drop banned source IPs IN-KERNEL via
+# cs-firewall-bouncer's nftables backend — the packet is dropped before it ever
+# reaches Traefik, costing zero per-request hops.
+#
+# Topology this respects (do NOT change without re-reading docs/architecture/networking.md):
+#   - Calico CNI + kube-proxy in IPTABLES mode (NOT eBPF).
+#   - Traefik is a LoadBalancer Service at 10.0.20.203, externalTrafficPolicy=Local
+#     (real client IP preserved — that's the whole point of the dedicated .203 IP).
+#   - LB traffic is DNAT'd to the Traefik POD, so the original source IP survives
+#     into the `forward` netfilter hook. The drop rule MUST therefore cover the
+#     `forward` hook, not only `input` (a pod-destined packet traverses forward,
+#     not input, on the node). Hence nftables_hooks: [input, forward] below.
+#
+# Packaging: cs-firewall-bouncer publishes NO container image. We pin the
+# official release binary (v0.0.34, 2025-08-04 — latest stable) and fetch it at
+# runtime: an initContainer (curlimages/curl — has curl + tar, alpine) downloads
+# + extracts the static binary into an emptyDir; the main container
+# (debian:bookworm-slim) runs it. The nftables backend talks netlink DIRECTLY
+# via github.com/google/nftables (go.mod + pkg/nftables/nftables.go: no os/exec)
+# and the docs confirm "mode nftables relies on github.com/google/nftables to
+# create table, chain and set" — so NO `nft` userspace CLI is needed and a plain
+# slim base image suffices. The binary is built CGO_ENABLED=0 / -extldflags
+# -static (Makefile), so it runs on glibc (debian) or musl (alpine) alike.
+#
+# Source: https://github.com/crowdsecurity/cs-firewall-bouncer
+#         https://docs.crowdsec.net/u/bouncers/firewall/
+#
+# nodeSelector pins this to ONE node (k8s-node2, which runs a Traefik pod) for first validation.
+# !!! REMOVING THE nodeSelector ROLLS THIS DAEMONSET CLUSTER-WIDE !!!
+# Do that ONLY after the one-node validation checklist passes (see commit/PR).
+
+locals {
+  # Pin a specific stable release. Bump deliberately (re-validate on one node first).
+  firewall_bouncer_version  = "v0.0.34"
+  firewall_bouncer_tgz_url  = "https://github.com/crowdsecurity/cs-firewall-bouncer/releases/download/${local.firewall_bouncer_version}/crowdsec-firewall-bouncer-linux-amd64.tgz"
+  firewall_bouncer_bin_path = "/opt/firewall-bouncer/crowdsec-firewall-bouncer"
+  firewall_bouncer_cfg_path = "/etc/crowdsec/bouncers/crowdsec-firewall-bouncer.yaml"
+
+  # Rendered firewall-bouncer config. Lives in a Secret (NOT a ConfigMap) because
+  # it embeds api_key. Key names/structure verified against the reference config
+  # (config/crowdsec-firewall-bouncer.yaml @ v0.0.34):
+  #   - `set-only` uses a HYPHEN (not set_only).
+  #   - `nftables_hooks` is a TOP-LEVEL list (sibling of `nftables:`, underscore).
+  #   - `deny_action` values are uppercase DROP / REJECT.
+  #   - `log_mode: stdout` sends logs to the container's stdout (default is `file`).
+  #   - api_url carries a trailing slash (matches the reference default).
+  firewall_bouncer_yaml = <<-YAML
+    mode: nftables
+    update_frequency: 10s
+    log_mode: stdout
+    log_level: info
+    api_url: http://crowdsec-service.crowdsec.svc.cluster.local:8080/
+    api_key: ${var.firewall_bouncer_key}
+    insecure_skip_verify: false
+    disable_ipv6: false
+    deny_action: DROP
+    deny_log: true
+    nftables:
+      ipv4:
+        enabled: true
+        set-only: false
+        table: crowdsec
+        chain: crowdsec-chain
+        priority: -10
+      ipv6:
+        enabled: true
+        set-only: false
+        table: crowdsec6
+        chain: crowdsec6-chain
+        priority: -10
+    nftables_hooks:
+      - input
+      - forward
+  YAML
+}
+
+resource "kubernetes_secret" "firewall_bouncer_config" {
+  metadata {
+    name      = "crowdsec-firewall-bouncer-config"
+    namespace = kubernetes_namespace.crowdsec.metadata[0].name
+    labels = {
+      "app.kubernetes.io/name" = "crowdsec-firewall-bouncer"
+      tier                     = var.tier
+    }
+    annotations = {
+      # Rotate the pods if the API key / config ever changes (the binary reads
+      # the config only at startup).
+      "reloader.stakater.com/match" = "true"
+    }
+  }
+  data = {
+    "crowdsec-firewall-bouncer.yaml" = local.firewall_bouncer_yaml
+  }
+  type = "Opaque"
+}
+
+resource "kubernetes_daemon_set_v1" "firewall_bouncer" {
+  metadata {
+    name      = "crowdsec-firewall-bouncer"
+    namespace = kubernetes_namespace.crowdsec.metadata[0].name
+    labels = {
+      "app.kubernetes.io/name" = "crowdsec-firewall-bouncer"
+      tier                     = var.tier
+    }
+  }
+  spec {
+    selector {
+      match_labels = {
+        "app.kubernetes.io/name" = "crowdsec-firewall-bouncer"
+      }
+    }
+    template {
+      metadata {
+        labels = {
+          "app.kubernetes.io/name" = "crowdsec-firewall-bouncer"
+          tier                     = var.tier
+        }
+        annotations = {
+          # Bounce pods when the config Secret changes (api_key rotation etc.).
+          "secret.reloader.stakater.com/reload" = kubernetes_secret.firewall_bouncer_config.metadata[0].name
+        }
+      }
+      spec {
+        priority_class_name = "tier-1-cluster"
+
+        # Program the HOST's nftables ruleset (not the pod netns) — the bouncer's
+        # drop rules must live in the host network namespace where DNAT'd LB
+        # traffic transits the forward hook.
+        host_network = true
+        dns_policy   = "ClusterFirstWithHostNet"
+
+        # ---- FIRST-VALIDATION PIN ----------------------------------------------
+        # Pinned to a SINGLE node so a mistake in the nftables rules can only
+        # affect one node. k8s-node2 is chosen because it currently runs a Traefik
+        # pod — required to validate the `forward`-hook drop on DNAT'd LoadBalancer
+        # traffic (under ETP=Local a node with no Traefik pod never sees that path,
+        # so the validation would be meaningless there).
+        # REMOVE this nodeSelector to roll the bouncer to EVERY node (the normal
+        # end state for a firewall bouncer) — but ONLY after the one-node
+        # validation checklist passes.
+        node_selector = {
+          "kubernetes.io/hostname" = "k8s-node2"
+        }
+        # ------------------------------------------------------------------------
+
+        # initContainer fetches + extracts the pinned release binary into the
+        # shared emptyDir. curlimages/curl is alpine and ships curl + tar.
+        init_container {
+          name  = "fetch-bouncer"
+          image = "curlimages/curl:8.10.1"
+          command = [
+            "sh", "-c",
+            <<-EOT
+              set -eu
+              echo "Downloading cs-firewall-bouncer ${local.firewall_bouncer_version}..."
+              curl -fsSL "${local.firewall_bouncer_tgz_url}" -o /tmp/fb.tgz
+              # Archive layout (verified @ v0.0.34): a single versioned top dir
+              # `crowdsec-firewall-bouncer-vX.Y.Z/` containing the binary plus
+              # config/, scripts/, install.sh. Strip that dir and extract ONLY the
+              # binary — the `*/crowdsec-firewall-bouncer` glob matches one path
+              # segment after the top dir, so config/...yaml is NOT pulled.
+              tar -xzf /tmp/fb.tgz -C /opt/firewall-bouncer --strip-components=1 \
+                --wildcards '*/crowdsec-firewall-bouncer'
+              chmod +x ${local.firewall_bouncer_bin_path}
+              echo "Fetched: $(ls -l ${local.firewall_bouncer_bin_path})"
+            EOT
+          ]
+          volume_mount {
+            name       = "binary"
+            mount_path = "/opt/firewall-bouncer"
+          }
+          resources {
+            requests = {
+              cpu    = "10m"
+              memory = "32Mi"
+            }
+            limits = {
+              memory = "64Mi"
+            }
+          }
+        }
+
+        container {
+          name  = "firewall-bouncer"
+          image = "debian:bookworm-slim"
+          command = [
+            local.firewall_bouncer_bin_path,
+            "-c", local.firewall_bouncer_cfg_path,
+          ]
+
+          # nftables backend needs NET_ADMIN to program the host ruleset. NET_RAW
+          # is the proven-safe companion the reference container images add. We
+          # deliberately AVOID full `privileged: true` — these two caps are
+          # sufficient for the netlink nftables path (no iptables/ipset shell-out
+          # here). If validation shows rules are NOT being installed, the next
+          # thing to try is privileged:true (see checklist) — but start minimal.
+          security_context {
+            capabilities {
+              add = ["NET_ADMIN", "NET_RAW"]
+            }
+          }
+
+          volume_mount {
+            name       = "binary"
+            mount_path = "/opt/firewall-bouncer"
+            read_only  = true
+          }
+          volume_mount {
+            name       = "config"
+            mount_path = local.firewall_bouncer_cfg_path
+            sub_path   = "crowdsec-firewall-bouncer.yaml"
+            read_only  = true
+          }
+
+          # No liveness probe: the bouncer runs as PID 1, so a crash — or a bad
+          # config that makes it exit non-zero at startup — surfaces on its own as
+          # a pod restart / CrashLoopBackOff. This avoids coupling pod liveness to
+          # a periodic LAPI round-trip (a brief LAPI blip must NOT bounce the pod).
+
+          resources {
+            requests = {
+              cpu    = "10m"
+              memory = "64Mi"
+            }
+            # crowdsec-quota enforces limits.memory on every pod in the ns.
+            limits = {
+              memory = "128Mi"
+            }
+          }
+        }
+
+        volume {
+          name = "binary"
+          empty_dir {}
+        }
+        volume {
+          name = "config"
+          secret {
+            secret_name = kubernetes_secret.firewall_bouncer_config.metadata[0].name
+          }
+        }
+      }
+    }
+  }
+  lifecycle {
+    # KYVERNO_LIFECYCLE_V1: Kyverno admission webhook mutates dns_config with ndots=2
+    ignore_changes = [spec[0].template[0].spec[0].dns_config]
+  }
+}
diff --git a/stacks/crowdsec/modules/crowdsec/main.tf b/stacks/crowdsec/modules/crowdsec/main.tf
index a900e8c4..86b8c3ab 100644
--- a/stacks/crowdsec/modules/crowdsec/main.tf
+++ b/stacks/crowdsec/modules/crowdsec/main.tf
@@ -21,6 +21,16 @@ variable "ingress_bouncer_key" {
   sensitive   = true
   description = "API key for the Traefik CrowdSec bouncer plugin. Seeded into LAPI via BOUNCER_KEY_traefik so the bouncer authenticates and pulls decisions — the same key the traefik-stack middleware presents."
 }
+variable "kvsync_bouncer_key" {
+  type        = string
+  sensitive   = true
+  description = "API key for the LAPI->Cloudflare-KV sync job (proxied-edge control plane). Seeded into LAPI via BOUNCER_KEY_kvsync; the rybbit-stack CronJob presents the same key to pull decisions."
+}
+variable "firewall_bouncer_key" {
+  type        = string
+  sensitive   = true
+  description = "API key for the cs-firewall-bouncer DaemonSet (direct-host in-kernel enforcement). Seeded into LAPI via BOUNCER_KEY_firewall; the DaemonSet presents the same key to stream decisions."
+}
 
 module "tls_secret" {
   source          = "../../../../modules/kubernetes/setup_tls_secret"
@@ -102,6 +112,15 @@ resource "kubernetes_config_map" "crowdsec_whitelist" {
         reason: "Trusted IP - never block"
         ip:
           - "176.12.22.76"
+        cidr:
+          # Never ban internal/cluster/LAN/tailnet sources. Enforcement (edge
+          # Worker + firewall-bouncer) drops on real source IP, so an internal
+          # range slipping into a decision could blackhole legit traffic — this
+          # makes that structurally impossible at the decision layer.
+          - "10.0.0.0/8"        # k8s nodes/pods/services + VLAN 10/20
+          - "172.16.0.0/12"     # RFC1918
+          - "192.168.0.0/16"    # LAN (192.168.1.0/24) + Sofia
+          - "100.64.0.0/10"     # Headscale tailnet (CGNAT)
       ---
       name: viktor/immich-asset-paths-whitelist
       description: "Don't penalise legit Immich timeline bursts (mobile scrub, web grid)"
@@ -153,7 +172,7 @@ resource "helm_release" "crowdsec" {
   repository = "https://crowdsecurity.github.io/helm-charts"
   chart      = "crowdsec"
 
-  values        = [templatefile("${path.module}/values.yaml", { homepage_username = var.homepage_username, homepage_password = var.homepage_password, DB_PASSWORD = var.db_password, ENROLL_KEY = var.enroll_key, SLACK_WEBHOOK_URL = var.slack_webhook_url, mysql_host = var.mysql_host, postgresql_host = var.postgresql_host, INGRESS_CROWDSEC_API_KEY = var.ingress_bouncer_key })]
+  values        = [templatefile("${path.module}/values.yaml", { homepage_username = var.homepage_username, homepage_password = var.homepage_password, DB_PASSWORD = var.db_password, ENROLL_KEY = var.enroll_key, SLACK_WEBHOOK_URL = var.slack_webhook_url, mysql_host = var.mysql_host, postgresql_host = var.postgresql_host, INGRESS_CROWDSEC_API_KEY = var.ingress_bouncer_key, KVSYNC_CROWDSEC_API_KEY = var.kvsync_bouncer_key, FIREWALL_CROWDSEC_API_KEY = var.firewall_bouncer_key })]
   timeout       = 1200
   wait          = true
   wait_for_jobs = true
diff --git a/stacks/crowdsec/modules/crowdsec/values.yaml b/stacks/crowdsec/modules/crowdsec/values.yaml
index 040b44d8..f03dfb22 100644
--- a/stacks/crowdsec/modules/crowdsec/values.yaml
+++ b/stacks/crowdsec/modules/crowdsec/values.yaml
@@ -135,6 +135,14 @@ lapi:
     # the prior manual registration was lost in the MySQL→PostgreSQL migration).
     - name: BOUNCER_KEY_traefik
       value: "${INGRESS_CROWDSEC_API_KEY}"
+    # Real enforcement path that replaces the dead Traefik Yaegi plugin:
+    #   kvsync   -> LAPI->Cloudflare-KV sync CronJob (proxied hosts, edge Worker)
+    #   firewall -> cs-firewall-bouncer DaemonSet (direct hosts, in-kernel nftables drop)
+    # Registered at LAPI startup (idempotent across the 3 replicas / restarts).
+    - name: BOUNCER_KEY_kvsync
+      value: "${KVSYNC_CROWDSEC_API_KEY}"
+    - name: BOUNCER_KEY_firewall
+      value: "${FIREWALL_CROWDSEC_API_KEY}"
   dashboard:
     enabled: true
     env:
diff --git a/stacks/rybbit/crowdsec_edge.tf b/stacks/rybbit/crowdsec_edge.tf
new file mode 100644
index 00000000..0c1ff5dc
--- /dev/null
+++ b/stacks/rybbit/crowdsec_edge.tf
@@ -0,0 +1,299 @@
+# =============================================================================
+# CrowdSec edge enforcement for Cloudflare-PROXIED hosts — control plane
+# =============================================================================
+# Proxied hosts terminate at the Cloudflare edge, so the in-cluster CrowdSec
+# bouncer (which keys on the real client IP seen by Traefik) never gets to
+# decide on them. To enforce CrowdSec bans/captchas on proxied traffic we push
+# the decision INTO the Cloudflare edge as account-level IP Lists + a single
+# zone-scoped WAF custom rule:
+#
+#   * Two account IP Lists — `crowdsec_ban` and `crowdsec_captcha` — hold the
+#     banned / captcha'd source IPs (empty in TF; populated at runtime).
+#   * A zone-scoped WAF ruleset in the http_request_firewall_custom phase
+#     blocks `(ip.src in $crowdsec_ban)` and managed-challenges
+#     `(ip.src in $crowdsec_captcha)`. Because it's a ZONE rule it enforces
+#     across ALL proxied hosts in the zone (~135), not just the handful a
+#     Worker would route. (The previous Worker+KV design only covered the ~27
+#     hosts the rybbit Worker routed; the analytics Worker in worker/ is
+#     unrelated and stays.)
+#
+# This file is the CONTROL PLANE that keeps those lists in sync with LAPI:
+#   1. the two empty IP Lists (list ITEMS are owned by the CronJob at runtime,
+#      NOT by Terraform — see the lifecycle ignore_changes on `item`),
+#   2. a LEAST-PRIVILEGE Cloudflare API token (account Filter-Lists edit only,
+#      scoped to this account) the sync job authenticates with,
+#   3. a CronJob running lapi_kv_sync.py every 2 min to full-reconcile LAPI
+#      decisions into the two lists (mirrors monitoring/alert_digest.tf: stock
+#      python:3.12-alpine + pure-stdlib script from a ConfigMap, no pip/apk at
+#      runtime).
+#
+# Cloudflare provider is pinned v4.52.7 (~> 4) — v4 schema is used throughout
+# (v5 differs greatly: policy is a block here not a `policies = [...]` list;
+# resources is a map not a jsonencode'd string; ruleset `rules` is a repeatable
+# block; list items use `item { value { ip = ... } }`; permission groups are
+# looked up via data.cloudflare_api_token_permission_groups, not a v5 *_list
+# data source). context7 only indexes v5, so the v4 arguments below were
+# verified against the v4.52.7 provider docs (github tag v4.52.7) — items
+# FLAGGED ### VERIFY for tg-plan are noted inline.
+# =============================================================================
+
+data "cloudflare_accounts" "main" {}
+
+locals {
+  cf_account_id = data.cloudflare_accounts.main.accounts[0].id
+}
+
+# -----------------------------------------------------------------------------
+# IP Lists — empty shells. The CronJob owns the items at runtime via the CF
+# Rules-Lists API; TF must NOT manage items or every 2-min sync would fight the
+# next `terragrunt apply` (apply would try to delete the runtime items).
+#
+# ### VERIFY (v4.52.7): cloudflare_list args account_id/name/kind/description;
+#     kind="ip" is one of {ip, redirect, hostname, asn}. The optional items
+#     block is named `item` (singular, Block Set) with `item { value { ip=... }
+#     comment=... }`. We declare NO `item` blocks (empty list) and
+#     ignore_changes=[item] so runtime items don't show as drift.
+#     NOTE: list `name` must match /^[a-zA-Z0-9_]+$/ (underscores ok, no dashes)
+#     — hence crowdsec_ban / crowdsec_captcha (underscore, not dash).
+# -----------------------------------------------------------------------------
+resource "cloudflare_list" "crowdsec_ban" {
+  account_id  = local.cf_account_id
+  name        = "crowdsec_ban"
+  kind        = "ip"
+  description = "CrowdSec ban decisions (synced from LAPI)"
+
+  lifecycle {
+    # The crowdsec-cf-sync CronJob adds/removes items at runtime; TF owns only
+    # the empty list shell. Without this, every apply would delete live bans.
+    ignore_changes = [item]
+  }
+}
+
+resource "cloudflare_list" "crowdsec_captcha" {
+  account_id  = local.cf_account_id
+  name        = "crowdsec_captcha"
+  kind        = "ip"
+  description = "CrowdSec captcha decisions (synced from LAPI)"
+
+  lifecycle {
+    ignore_changes = [item]
+  }
+}
+
+# -----------------------------------------------------------------------------
+# Zone-scoped WAF custom ruleset — the actual enforcement. One ruleset, two
+# rules, applied to EVERY proxied host in the zone.
+#
+# ### VERIFY (v4.52.7): cloudflare_ruleset with zone_id + kind="zone" +
+#     phase="http_request_firewall_custom"; `rules` is a repeatable block with
+#     action/expression/description/enabled. actions "block" and
+#     "managed_challenge" are both valid. List references in WAF expressions use
+#     the list NAME with a `$` prefix (NOT the list id): ($crowdsec_ban).
+#     Rule order matters — ban (block) is evaluated before captcha so a
+#     double-listed IP is blocked outright (the sync script also enforces
+#     ban-wins, so an IP is never in both lists, but order is belt-and-braces).
+#
+# zone_id is the viktorbarzin.me zone — the single zone id used repo-wide
+# (default of var.cloudflare_zone_id in modules/kubernetes/ingress_factory and
+# hardcoded the same in stacks/kms/main.tf; source of truth is the git-crypt'd
+# config.tfvars). Hardcoded here (with the conventional marker comment) because
+# the rybbit stack does not import the ingress_factory module.
+# -----------------------------------------------------------------------------
+resource "cloudflare_ruleset" "crowdsec" {
+  zone_id     = "fd2c5dd4efe8fe38958944e74d0ced6d" # cloudflare_zone_id (viktorbarzin.me)
+  name        = "crowdsec-ip-enforcement"
+  description = "Block/challenge IPs CrowdSec flagged (synced from LAPI into CF IP Lists)"
+  kind        = "zone"
+  phase       = "http_request_firewall_custom"
+
+  rules {
+    action      = "block"
+    expression  = "(ip.src in $crowdsec_ban)"
+    description = "CrowdSec: block banned IPs"
+    enabled     = true
+  }
+
+  rules {
+    action      = "managed_challenge"
+    expression  = "(ip.src in $crowdsec_captcha)"
+    description = "CrowdSec: managed-challenge captcha'd IPs"
+    enabled     = true
+  }
+}
+
+# -----------------------------------------------------------------------------
+# Least-privilege API token for the sync job: account-level Filter-Lists edit
+# ONLY, scoped to this single account (no zone/DNS/Workers access). The token
+# value is sensitive and lands in TF state (Tier-1 PG, encrypted at rest) and
+# in the rybbit Secret below — same trust level as the CF Global API Key
+# already in state.
+#
+# ### VERIFY (v4.52.7): cloudflare_api_token with a repeatable `policy` block
+#     (effect / permission_groups = Set of String / resources = Map of String);
+#     token secret is exposed as `.value` (sensitive).
+#
+# ### VERIFY — PERMISSION GROUP NAME (highest-risk item). v4.52.7 deprecates
+#     the flat `.permissions[...]` map ("some permissions overlap resource
+#     scope"); the non-deprecated lookup is the scoped `.account[...]` map.
+#     Cloudflare's current permissions reference calls the account list-edit
+#     group "Account Filter Lists Edit" (and read "Account Filter Lists Read").
+#     An OLDER community gist instead shows "Account Rule Lists Read/Write" —
+#     Cloudflare has renamed this group over time. If `tg plan` errors with a
+#     missing key, try (in order): .account["Account Filter Lists Edit"] ->
+#     .account["Account Rule Lists Write"], or enumerate the live names with:
+#       terraform console
+#       > data.cloudflare_api_token_permission_groups.all.account
+#     Read is not strictly required for edit (Edit = full CRUDL) but the sync
+#     job GETs items, so we include Read too to be safe.
+# -----------------------------------------------------------------------------
+data "cloudflare_api_token_permission_groups" "all" {}
+
+resource "cloudflare_api_token" "list_sync" {
+  name = "rybbit-crowdsec-list-sync"
+
+  policy {
+    effect = "allow"
+    permission_groups = [
+      data.cloudflare_api_token_permission_groups.all.account["Account Rule Lists Write"],
+      data.cloudflare_api_token_permission_groups.all.account["Account Rule Lists Read"],
+    ]
+    resources = {
+      "com.cloudflare.api.account.${local.cf_account_id}" = "*"
+    }
+  }
+}
+
+# -----------------------------------------------------------------------------
+# Pure-stdlib sync script, mounted into the CronJob from a ConfigMap (the
+# alert_digest pattern — no per-run package installs).
+# -----------------------------------------------------------------------------
+resource "kubernetes_config_map" "crowdsec_cf_sync_script" {
+  metadata {
+    name      = "crowdsec-cf-sync-script"
+    namespace = "rybbit"
+  }
+  data = {
+    "lapi_kv_sync.py" = file("${path.module}/lapi_kv_sync.py")
+  }
+}
+
+# Secrets consumed by the sync job: the LAPI bouncer key (registered in LAPI,
+# stored in Vault secret/platform -> kvsync_bouncer_key) and the minted CF
+# token value. Account id and list ids are NOT secret and are passed as plain
+# env values on the CronJob.
+resource "kubernetes_secret" "crowdsec_cf_sync" {
+  metadata {
+    name      = "crowdsec-cf-sync"
+    namespace = "rybbit"
+  }
+  type = "Opaque"
+  data = {
+    LAPI_KEY     = data.vault_kv_secret_v2.cf_platform.data["kvsync_bouncer_key"]
+    CF_API_TOKEN = cloudflare_api_token.list_sync.value
+  }
+}
+
+resource "kubernetes_cron_job_v1" "crowdsec_cf_sync" {
+  metadata {
+    name      = "crowdsec-cf-sync"
+    namespace = "rybbit"
+    labels = {
+      app  = "crowdsec-cf-sync"
+      tier = local.tiers.aux
+    }
+  }
+  spec {
+    concurrency_policy            = "Forbid"
+    failed_jobs_history_limit     = 3
+    successful_jobs_history_limit = 3
+    schedule                      = "*/2 * * * *"
+    starting_deadline_seconds     = 110
+    job_template {
+      metadata {}
+      spec {
+        backoff_limit              = 2
+        ttl_seconds_after_finished = 3600
+        template {
+          metadata {
+            labels = {
+              app = "crowdsec-cf-sync"
+            }
+          }
+          spec {
+            restart_policy = "OnFailure"
+            container {
+              name              = "crowdsec-cf-sync"
+              image             = "docker.io/library/python:3.12-alpine"
+              image_pull_policy = "IfNotPresent"
+              command           = ["python3", "/scripts/lapi_kv_sync.py"]
+              env {
+                name = "LAPI_KEY"
+                value_from {
+                  secret_key_ref {
+                    name = kubernetes_secret.crowdsec_cf_sync.metadata[0].name
+                    key  = "LAPI_KEY"
+                  }
+                }
+              }
+              env {
+                name = "CF_API_TOKEN"
+                value_from {
+                  secret_key_ref {
+                    name = kubernetes_secret.crowdsec_cf_sync.metadata[0].name
+                    key  = "CF_API_TOKEN"
+                  }
+                }
+              }
+              env {
+                name  = "CF_ACCOUNT_ID"
+                value = local.cf_account_id
+              }
+              env {
+                name  = "CF_BAN_LIST_ID"
+                value = cloudflare_list.crowdsec_ban.id
+              }
+              env {
+                name  = "CF_CAPTCHA_LIST_ID"
+                value = cloudflare_list.crowdsec_captcha.id
+              }
+              env {
+                name  = "PUSHGATEWAY_URL"
+                value = "http://prometheus-prometheus-pushgateway.monitoring:9091"
+              }
+              volume_mount {
+                name       = "script"
+                mount_path = "/scripts"
+                read_only  = true
+              }
+              resources {
+                requests = {
+                  cpu    = "10m"
+                  memory = "48Mi"
+                }
+                limits = {
+                  memory = "96Mi"
+                }
+              }
+            }
+            volume {
+              name = "script"
+              config_map {
+                name = kubernetes_config_map.crowdsec_cf_sync_script.metadata[0].name
+              }
+            }
+            dns_config {
+              option {
+                name  = "ndots"
+                value = "2"
+              }
+            }
+          }
+        }
+      }
+    }
+  }
+  lifecycle {
+    # KYVERNO_LIFECYCLE_V1: Kyverno admission webhook mutates dns_config with ndots=2
+    ignore_changes = [spec[0].job_template[0].spec[0].template[0].spec[0].dns_config]
+  }
+}
diff --git a/stacks/rybbit/lapi_kv_sync.py b/stacks/rybbit/lapi_kv_sync.py
new file mode 100644
index 00000000..942912c9
--- /dev/null
+++ b/stacks/rybbit/lapi_kv_sync.py
@@ -0,0 +1,318 @@
+#!/usr/bin/env python3
+"""Sync CrowdSec LAPI decisions -> two Cloudflare account IP Lists.
+
+Cloudflare-PROXIED hosts terminate at the CF edge, so the in-cluster CrowdSec
+bouncer (which keys on the client IP Traefik sees) never decides on them. We
+push the decisions into the edge instead: a zone-scoped WAF custom rule blocks
+`(ip.src in $crowdsec_ban)` and managed-challenges `(ip.src in $crowdsec_captcha)`
+across EVERY proxied host in the zone. This job is the control plane that keeps
+those two IP Lists in sync with LAPI.
+
+(Filename kept as lapi_kv_sync.py for path/ConfigMap continuity with the prior
+Workers-KV design; it no longer touches KV — it reconciles CF Rules Lists.)
+
+Design notes:
+  * Pure Python stdlib (no pip/apk at runtime) — runs on stock python:3.12-alpine
+    mounted from a ConfigMap, the alert_digest pattern.
+  * FULL RECONCILE each run: read the complete decision set from LAPI, partition
+    into ban / captcha desired sets, then for each list compute add (desired -
+    existing) and remove (existing - desired) and apply both. An IP listed for
+    BOTH ban and captcha is placed in BAN ONLY (ban wins; the WAF rule order
+    also blocks-before-challenges as belt-and-braces). A `cscli decisions
+    delete` therefore clears from the edge within one interval (<=2 min).
+  * FAIL-SAFE on LAPI: if LAPI can't be read we SKIP the run (lists untouched,
+    exit 0). A LAPI outage thus freezes the edge state rather than wiping every
+    ban — degrade toward the last-known-good block set, never toward all-block
+    or a thundering un-ban. (Decisions linger only until the next successful
+    sync, not their TTL — we reconcile to LAPI truth, we don't expire entries.)
+  * FAIL-LOUD on Cloudflare: any CF API error is logged and the job exits
+    non-zero so the failure is visible (CronJob backoff + missing success
+    metric + the next run retries).
+
+Cloudflare Rules-Lists API (account-level IP list items), verified against the
+official API reference (developers.cloudflare.com, 2026):
+  * GET    /accounts/{acct}/rules/lists/{list}/items   -> paginated; next page
+           cursor at result_info.cursors.after, passed back as ?cursor=. Each
+           item = {"id","ip","created_on",...}.
+  * POST   /accounts/{acct}/rules/lists/{list}/items   -> body JSON ARRAY
+           [{"ip":"1.2.3.4"},...]. APPENDS/upserts (does NOT replace the list).
+           ASYNCHRONOUS: returns {"result":{"operation_id":...}}.
+  * DELETE /accounts/{acct}/rules/lists/{list}/items   -> body {"items":[{"id":
+           "<item_id>"},...]} (delete by item id, not ip). ASYNCHRONOUS.
+  * GET    /accounts/{acct}/rules/lists/bulk_operations/{op_id} -> status in
+           {pending,running,completed,failed} (failed carries `error`).
+  ASYNC HANDLING: Cloudflare allows only ONE pending bulk operation per ACCOUNT.
+  So we must NOT fire add+delete (or both lists) concurrently — we serialize and
+  poll each operation_id to a terminal state (short bounded timeout) before the
+  next mutation. If a poll times out we stop mutating for this run and report
+  partial success (the next 2-min run reconciles the rest); we never abandon an
+  in-flight op and immediately issue another (that would 409/reject).
+"""
+import json
+import os
+import sys
+import time
+import urllib.error
+import urllib.parse
+import urllib.request
+
+LAPI_URL = os.environ.get(
+    "LAPI_URL", "http://crowdsec-service.crowdsec.svc.cluster.local:8080"
+).rstrip("/")
+LAPI_KEY = os.environ["LAPI_KEY"]  # kvsync bouncer key, registered in LAPI
+CF_ACCOUNT_ID = os.environ["CF_ACCOUNT_ID"]
+CF_API_TOKEN = os.environ["CF_API_TOKEN"]  # scoped: Account Filter Lists Edit
+CF_BAN_LIST_ID = os.environ["CF_BAN_LIST_ID"]
+CF_CAPTCHA_LIST_ID = os.environ["CF_CAPTCHA_LIST_ID"]
+PUSHGATEWAY = os.environ.get("PUSHGATEWAY_URL", "").rstrip("/")  # optional
+
+CF_API = "https://api.cloudflare.com/client/v4"
+# Cloudflare item objects expose the ip differently between list and create
+# responses; for an IP-kind list each GET item carries a top-level "ip".
+# Batch sizes: no official per-request cap is documented, so keep batches
+# generous but bounded (well under the global 1200 req / 5 min limit).
+BATCH = 1000
+# Async op polling: 1 pending bulk op per account, so poll to terminal state.
+POLL_TIMEOUT = 25  # seconds to wait for one bulk op (the run has ~110s budget)
+POLL_INTERVAL = 1.0
+
+
+class CFError(Exception):
+    """Cloudflare API failure -> job should exit non-zero (fail loud)."""
+
+
+def _req(url, *, method="GET", headers=None, data=None, timeout=20):
+    req = urllib.request.Request(url, method=method, headers=headers or {}, data=data)
+    with urllib.request.urlopen(req, timeout=timeout) as resp:
+        body = resp.read()
+    return json.loads(body) if body else None
+
+
+def _cf(url, *, method="GET", payload=None, timeout=20):
+    """Call the CF API with the bearer token; raise CFError on any failure."""
+    headers = {"Authorization": f"Bearer {CF_API_TOKEN}"}
+    data = None
+    if payload is not None:
+        headers["Content-Type"] = "application/json"
+        data = json.dumps(payload).encode()
+    try:
+        res = _req(url, method=method, headers=headers, data=data, timeout=timeout)
+    except urllib.error.HTTPError as e:
+        detail = ""
+        try:
+            detail = e.read().decode(errors="replace")[:500]
+        except Exception:
+            pass
+        raise CFError(f"{method} {url} -> HTTP {e.code} {detail}") from e
+    except urllib.error.URLError as e:
+        raise CFError(f"{method} {url} -> {e}") from e
+    if res is not None and not res.get("success", True):
+        raise CFError(f"{method} {url} -> not success: {res.get('errors')}")
+    return res
+
+
+# --------------------------------------------------------------------------- #
+# LAPI
+# --------------------------------------------------------------------------- #
+def fetch_decisions():
+    """Return (ban_set, captcha_set) of IPs from LAPI.
+
+    Only scope=="ip" decisions are projected (the WAF rule keys on ip.src). An
+    IP appearing in both ban and captcha is placed in BAN only. Raises on
+    transport/HTTP error so the caller can SKIP the run (fail-safe).
+    """
+    data = _req(
+        f"{LAPI_URL}/v1/decisions",
+        headers={"X-Api-Key": LAPI_KEY, "Accept": "application/json"},
+    )
+    ban, captcha = set(), set()
+    for d in data or []:
+        if (d.get("scope") or "").lower() != "ip":
+            continue
+        ip = d.get("value")
+        if not ip:
+            continue
+        dtype = (d.get("type") or "").lower()
+        if dtype == "ban":
+            ban.add(ip)
+        elif dtype == "captcha":
+            captcha.add(ip)
+        # other remediation types (e.g. throttle) are ignored
+    captcha -= ban  # ban wins: never list the same IP in both
+    return ban, captcha
+
+
+# --------------------------------------------------------------------------- #
+# Cloudflare list items
+# --------------------------------------------------------------------------- #
+def cf_list_items(list_id):
+    """Return {ip: item_id} for every item currently in the list (paginated)."""
+    out = {}
+    cursor = ""
+    while True:
+        url = f"{CF_API}/accounts/{CF_ACCOUNT_ID}/rules/lists/{list_id}/items?per_page=1000"
+        if cursor:
+            url += f"&cursor={urllib.parse.quote(cursor)}"
+        res = _cf(url)
+        for it in (res.get("result") or []):
+            ip = it.get("ip")
+            if ip:
+                out[ip] = it.get("id")
+        cursor = (((res.get("result_info") or {}).get("cursors") or {}).get("after")) or ""
+        if not cursor:
+            return out
+
+
+def _wait_for_op(op_id):
+    """Poll a bulk operation to a terminal state. Returns True if completed,
+    False if it timed out (still pending/running). Raises CFError if it failed.
+
+    We must reach a terminal state before issuing the next mutation: CF allows
+    only one pending bulk op per account, so firing another while this is
+    in-flight would be rejected.
+    """
+    if not op_id:
+        return True
+    deadline = time.time() + POLL_TIMEOUT
+    url = f"{CF_API}/accounts/{CF_ACCOUNT_ID}/rules/lists/bulk_operations/{op_id}"
+    while time.time() < deadline:
+        res = _cf(url)
+        status = ((res.get("result") or {}).get("status") or "").lower()
+        if status == "completed":
+            return True
+        if status == "failed":
+            raise CFError(f"bulk op {op_id} failed: {(res.get('result') or {}).get('error')}")
+        time.sleep(POLL_INTERVAL)
+    print(f"[warn] bulk op {op_id} still {status or 'pending'} after {POLL_TIMEOUT}s; "
+          f"stopping further mutations this run (next run reconciles)", file=sys.stderr)
+    return False
+
+
+def _op_id(res):
+    return ((res or {}).get("result") or {}).get("operation_id")
+
+
+def cf_add_items(list_id, ips):
+    """POST new IPs to the list (append). Returns the operation_id (async)."""
+    # If callers ever exceed one batch, each POST is a separate bulk op and the
+    # single-pending-op rule forces us to wait between them.
+    last_op = None
+    for i in range(0, len(ips), BATCH):
+        chunk = [{"ip": ip} for ip in ips[i : i + BATCH]]
+        res = _cf(
+            f"{CF_API}/accounts/{CF_ACCOUNT_ID}/rules/lists/{list_id}/items",
+            method="POST",
+            payload=chunk,
+        )
+        last_op = _op_id(res)
+        if i + BATCH < len(ips):  # more chunks coming -> serialize
+            if not _wait_for_op(last_op):
+                return last_op  # timed out; bail (partial), next run continues
+    return last_op
+
+
+def cf_delete_items(list_id, item_ids):
+    """DELETE items by id. Returns the operation_id (async)."""
+    last_op = None
+    for i in range(0, len(item_ids), BATCH):
+        chunk = {"items": [{"id": iid} for iid in item_ids[i : i + BATCH]]}
+        res = _cf(
+            f"{CF_API}/accounts/{CF_ACCOUNT_ID}/rules/lists/{list_id}/items",
+            method="DELETE",
+            payload=chunk,
+        )
+        last_op = _op_id(res)
+        if i + BATCH < len(item_ids):
+            if not _wait_for_op(last_op):
+                return last_op
+    return last_op
+
+
+def reconcile(label, list_id, desired):
+    """Reconcile one list to `desired` (a set of IPs).
+
+    Returns (added, removed). Serializes add->wait->delete so we respect the
+    one-pending-bulk-op-per-account limit. Raises CFError on hard failure.
+    """
+    existing = cf_list_items(list_id)  # {ip: item_id}
+    existing_ips = set(existing)
+    to_add = sorted(desired - existing_ips)
+    to_remove_ids = [existing[ip] for ip in (existing_ips - desired)]
+
+    added = removed = 0
+    if to_add:
+        op = cf_add_items(list_id, to_add)
+        if _wait_for_op(op):
+            added = len(to_add)
+        else:
+            # add op didn't finish; skip the delete this run to avoid stacking a
+            # second pending op, and report what we attempted.
+            print(f"[warn] {label}: add op not confirmed; deferring deletes", file=sys.stderr)
+            return added, removed
+    if to_remove_ids:
+        op = cf_delete_items(list_id, to_remove_ids)
+        if _wait_for_op(op):
+            removed = len(to_remove_ids)
+    print(f"[ok] {label}: +{added} / -{removed} (desired={len(desired)}, was={len(existing_ips)})")
+    return added, removed
+
+
+# --------------------------------------------------------------------------- #
+# Metrics (best-effort)
+# --------------------------------------------------------------------------- #
+def push_metrics(ban_n, captcha_n, ok):
+    if not PUSHGATEWAY:
+        return
+    payload = (
+        "# TYPE crowdsec_cf_list_ban_count gauge\n"
+        f"crowdsec_cf_list_ban_count {ban_n}\n"
+        "# TYPE crowdsec_cf_list_captcha_count gauge\n"
+        f"crowdsec_cf_list_captcha_count {captcha_n}\n"
+        "# TYPE crowdsec_cf_list_sync_success gauge\n"
+        f"crowdsec_cf_list_sync_success {1 if ok else 0}\n"
+        "# TYPE crowdsec_cf_list_sync_last_run_seconds gauge\n"
+        f"crowdsec_cf_list_sync_last_run_seconds {int(time.time())}\n"
+    )
+    try:
+        _req(
+            f"{PUSHGATEWAY}/metrics/job/crowdsec-cf-list-sync",
+            method="PUT",
+            headers={"Content-Type": "text/plain"},
+            data=payload.encode(),
+            timeout=10,
+        )
+    except Exception as e:  # metrics are best-effort, never fail the job
+        print(f"[warn] pushgateway: {e}", file=sys.stderr)
+
+
+# --------------------------------------------------------------------------- #
+def main():
+    # 1. Desired state from LAPI. Any failure here = SKIP (fail-safe).
+    try:
+        ban, captcha = fetch_decisions()
+    except Exception as e:
+        print(
+            f"[skip] LAPI unreadable ({e}); leaving CF lists untouched "
+            f"(fail-safe: freeze last-known edge state).",
+            file=sys.stderr,
+        )
+        push_metrics(0, 0, ok=False)
+        return 0
+
+    print(f"[info] LAPI desired: {len(ban)} ban / {len(captcha)} captcha (ip-scope)")
+
+    # 2. Reconcile both lists. CF errors fail loud (non-zero exit).
+    try:
+        reconcile("ban", CF_BAN_LIST_ID, ban)
+        reconcile("captcha", CF_CAPTCHA_LIST_ID, captcha)
+    except CFError as e:
+        print(f"[error] Cloudflare API failure: {e}", file=sys.stderr)
+        push_metrics(len(ban), len(captcha), ok=False)
+        return 1
+
+    push_metrics(len(ban), len(captcha), ok=True)
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())