diff --git a/modules/ai-guardrails/README.md b/modules/ai-guardrails/README.md
new file mode 100644
index 0000000..4c0c9f7
--- /dev/null
+++ b/modules/ai-guardrails/README.md
@@ -0,0 +1,96 @@
+# ai-guardrails
+
+Org-policy guardrails for AI workloads — the prevention half of this repo's AI
+story. The risks [CloudCanary](https://github.com/ChrisInvictus/CloudCanary)
+*detects* become, where org policy can reach them, structurally impossible.
+Companion to the [MCP trust-boundary threat model](../../docs/threat-models/mcp-trust-boundaries.md).
+
+Every constraint this module applies is a **Google-managed (canned) org policy
+constraint, verified present against a live organization** before the module was
+written. No invented constraint names; the module's claims are checkable against
+`GET /v2/{parent}/constraints`.
+
+## The loop this closes
+
+| Risk | Threat model | CloudCanary detects | This module prevents |
+|---|---|---|---|
+| Shadow AI — Vertex/Gemini enabled where unsanctioned | B2/B3, LLM03 | `ai-apis` adoption signal | `gcp.restrictServiceUsage` denies the AI surface at that scope |
+| Internet-facing AI development surface | B3 (exposure) | *(compute/firewall signals, indirectly)* | `ainotebooks.restrictPublicIp` |
+| Model data in unapproved regions | B3 (custody) | — | `gcp.resourceLocations` (opt-in, explicit regions) |
+| AI data under Google-managed keys only | B3 (custody) | — | `gcp.restrictNonCmekServices` (opt-in) |
+
+The adoption row is the pairing at its clearest: at scopes where AI is
+sanctioned, CloudCanary's signal is a welcome heads-up; at scopes where this
+module denies the services, the same event is an *error at enablement time* and
+the alert should never fire. Guardrails stop what you've predicted; the canary
+catches what you haven't.
+
+## What this module deliberately does not do
+
+- **Restrict who holds `roles/aiplatform.admin`.** Bounding role-holding is IAM
+  deny-policy territory, not org policy. Until an IAM deny policy is warranted,
+  CloudCanary's `ai-iam` signal is the compensating **detective** control — every
+  `aiplatform.*` grant is drift-alerted, so the allowlist question is answered by
+  review rather than silently unenforced. Named here so the gap is a decision,
+  not an omission.
+- **See model inference.** Prompt injection and tool-output poisoning (threat
+  model **B5**) happen inside the host application's reasoning loop. No
+  infrastructure constraint observes them — which is the threat model's central
+  thesis about what is genuinely new here.
+
+## Inputs
+
+| Name | Type | Default | Purpose |
+|---|---|---|---|
+| `target_id` | string | — | Policy parent: `organizations/NNN`, `folders/NNN`, `projects/ID` |
+| `enforce_restrict_ai_services` | bool | `false` | Deny AI services at this scope (turn on where AI is unsanctioned) |
+| `denied_ai_services` | list | aiplatform, generativelanguage, notebooks | The denied surface |
+| `enforce_no_public_notebooks` | bool | `true` | No external IPs on Workbench/notebooks |
+| `allowed_locations` | list | `[]` | AI resource residency (empty = not applied) |
+| `enforce_require_cmek` | bool | `false` | CMEK requirement for regulated estates |
+| `cmek_required_services` | list | aiplatform, notebooks | Services under the CMEK rule |
+
+## Outputs
+
+`guardrails_applied` — the enforced configuration, for audit and evidence.
+
+## Usage
+
+```hcl
+# Deny AI in the sandbox — experimentation happens on sanctioned paths only.
+module "sandbox_ai_guardrails" {
+  source    = "../ai-guardrails"
+  target_id = "folders/${var.sandbox_folder_id}"
+
+  enforce_restrict_ai_services = true # AI not sanctioned here
+  # no-public-notebooks defaults on; residency and CMEK stay off
+}
+
+# Govern AI in /workloads — allowed, but never on a public notebook,
+# and only in US regions.
+module "workloads_ai_guardrails" {
+  source    = "../ai-guardrails"
+  target_id = "folders/${var.workloads_folder_id}"
+
+  allowed_locations = ["in:us-locations"]
+}
+```
+
+Applied at a folder, a project vended by the
+[project factory](../project-factory/) into it is born with these guardrails the
+same way it's born with no default network — one more thing that's true before
+the first API call.
+
+## Design notes
+
+- **Deny-list, not allowlist, for services.** `gcp.restrictServiceUsage` here
+  denies the AI surface specifically; this module governs AI adoption, not the
+  whole service catalog. An estate-wide service allowlist is a bigger decision
+  that belongs to the org baseline, not an AI module.
+- **Secure-by-default where the default is free; opt-in where it costs.**
+  No-public-notebooks defaults on (no operational cost). Residency and CMEK are
+  explicit choices because they constrain real workflows and add key lifecycle.
+- **Roadmap.** The `ainotebooks.*` family has seven more constraints (root
+  access, file downloads, terminal, VPC restrictions, upgrade schedules) —
+  candidates as the AI surface here grows. Custom constraints (CEL) for
+  Vertex endpoint IAM are the next step beyond canned coverage.
diff --git a/modules/ai-guardrails/main.tf b/modules/ai-guardrails/main.tf
new file mode 100644
index 0000000..218be87
--- /dev/null
+++ b/modules/ai-guardrails/main.tf
@@ -0,0 +1,127 @@
+# AI guardrails — the prevention half of this repo's AI story.
+#
+# CloudCanary's WATCH_AI_WORKLOADS *detects* AI-workload risk signals; this
+# module makes the preventable ones impossible — the same prevent/detect pairing
+# the whole project is built on, applied to the newest workload class. Companion
+# to docs/threat-models/mcp-trust-boundaries.md.
+#
+# Every constraint below is a Google-managed (canned) org policy constraint,
+# verified present against a live organization before this module was written —
+# constraint names are ground truth, not plausible guesses:
+#
+#   gcp.restrictServiceUsage       deny AI services where they are not
+#                                  sanctioned — the guardrail form of
+#                                  CloudCanary's "AI API enabled" adoption
+#                                  signal. Shadow AI becomes unrepresentable
+#                                  at scopes where it is denied.
+#                                  → threat model B2/B3, OWASP LLM03.
+#
+#   ainotebooks.restrictPublicIp   no internet-facing Workbench/notebook
+#                                  instances — the AI-development surface's
+#                                  version of compute.vmExternalIpAccess.
+#                                  → threat model B3 (exposure).
+#
+#   gcp.resourceLocations          AI resources only in approved regions —
+#                                  where model artifacts and training data
+#                                  may live. → threat model B3 (custody).
+#
+#   gcp.restrictNonCmekServices    optional CMEK requirement for regulated
+#                                  estates; off by default at demo scale,
+#                                  same documented-and-deferred stance the
+#                                  factory takes on registry CMEK.
+#                                  → threat model B3 (custody).
+#
+# What org policy deliberately does NOT cover here — named, not omitted:
+#   - Restricting who may hold roles/aiplatform.admin is IAM-deny territory,
+#     not org policy. Until an IAM deny policy is warranted, CloudCanary's
+#     ai-iam signal is the compensating detective control (see README).
+#   - Model *inference* risks — prompt injection, tool-output poisoning
+#     (threat model B5) — live in the host application. No infrastructure
+#     constraint sees them; the threat model's thesis is precisely that this
+#     boundary is the new one.
+
+terraform {
+  required_version = ">= 1.7"
+  required_providers {
+    google = {
+      source  = "hashicorp/google"
+      version = "~> 6.0"
+    }
+  }
+}
+
+# ---------------------------------------------------------- shadow-AI prevention
+# Deny the AI service surface at scopes where AI work is not sanctioned.
+# Applied to /sandbox (or any non-AI folder), enabling Vertex there stops being
+# an alert and becomes an error. Deliberately a DENY list of AI services, not an
+# allowlist of everything else — this module governs AI adoption, not the whole
+# service catalog.
+resource "google_org_policy_policy" "restrict_ai_services" {
+  count = var.enforce_restrict_ai_services ? 1 : 0
+
+  name   = "${var.target_id}/policies/gcp.restrictServiceUsage"
+  parent = var.target_id
+
+  spec {
+    rules {
+      values {
+        denied_values = var.denied_ai_services
+      }
+    }
+  }
+}
+
+# ------------------------------------------------------- no public AI notebooks
+# Workbench/notebook instances are the AI-development surface most likely to
+# hold credentials, training data, and model access in one place. Boolean
+# constraint: no external IPs on them, ever, at this scope.
+resource "google_org_policy_policy" "no_public_ai_notebooks" {
+  count = var.enforce_no_public_notebooks ? 1 : 0
+
+  name   = "${var.target_id}/policies/ainotebooks.restrictPublicIp"
+  parent = var.target_id
+
+  spec {
+    rules {
+      enforce = "TRUE"
+    }
+  }
+}
+
+# ---------------------------------------------------------------- data residency
+# Where AI resources may physically live. Empty list = constraint not applied
+# (count gate), so residency is opt-in with explicit regions — the operator
+# states the policy rather than inheriting a guess.
+resource "google_org_policy_policy" "ai_resource_locations" {
+  count = length(var.allowed_locations) > 0 ? 1 : 0
+
+  name   = "${var.target_id}/policies/gcp.resourceLocations"
+  parent = var.target_id
+
+  spec {
+    rules {
+      values {
+        allowed_values = var.allowed_locations
+      }
+    }
+  }
+}
+
+# ------------------------------------------------------------------ CMEK (opt-in)
+# Require customer-managed keys for the listed AI services. Off by default:
+# CMEK adds a KMS keyring + rotation lifecycle that is pure operational surface
+# until there is regulated data to protect.
+resource "google_org_policy_policy" "require_ai_cmek" {
+  count = var.enforce_require_cmek ? 1 : 0
+
+  name   = "${var.target_id}/policies/gcp.restrictNonCmekServices"
+  parent = var.target_id
+
+  spec {
+    rules {
+      values {
+        denied_values = var.cmek_required_services
+      }
+    }
+  }
+}
diff --git a/modules/ai-guardrails/outputs.tf b/modules/ai-guardrails/outputs.tf
new file mode 100644
index 0000000..aa90071
--- /dev/null
+++ b/modules/ai-guardrails/outputs.tf
@@ -0,0 +1,10 @@
+output "guardrails_applied" {
+  description = "The enforced AI-guardrail configuration at this target, for audit and evidence"
+  value = {
+    target               = var.target_id
+    restrict_ai_services = var.enforce_restrict_ai_services ? var.denied_ai_services : []
+    no_public_notebooks  = var.enforce_no_public_notebooks
+    allowed_locations    = var.allowed_locations
+    require_cmek         = var.enforce_require_cmek ? var.cmek_required_services : []
+  }
+}
diff --git a/modules/ai-guardrails/variables.tf b/modules/ai-guardrails/variables.tf
new file mode 100644
index 0000000..062f3dd
--- /dev/null
+++ b/modules/ai-guardrails/variables.tf
@@ -0,0 +1,64 @@
+variable "target_id" {
+  description = <<-EOT
+    Where these guardrails apply, as an org-policy parent — "organizations/NNN",
+    "folders/NNN", or "projects/ID". Applied at a folder (e.g. /sandbox to deny
+    AI there, or /workloads to govern it), the constraints inherit downward like
+    the rest of the org baseline; the project factory can pass a vended project
+    for per-project opt-in.
+  EOT
+  type        = string
+}
+
+variable "enforce_restrict_ai_services" {
+  description = "Deny the AI service surface at this scope (shadow-AI prevention). Turn on where AI work is NOT sanctioned; leave off where it is."
+  type        = bool
+  default     = false
+}
+
+variable "denied_ai_services" {
+  description = "AI services denied when enforce_restrict_ai_services is on. gcp.restrictServiceUsage expects service names."
+  type        = list(string)
+  default = [
+    "aiplatform.googleapis.com",
+    "generativelanguage.googleapis.com",
+    "notebooks.googleapis.com",
+  ]
+}
+
+variable "enforce_no_public_notebooks" {
+  description = "Deny external IPs on Vertex Workbench / notebook instances (ainotebooks.restrictPublicIp). On by default — the AI-development surface's version of the org's no-public-VM stance."
+  type        = bool
+  default     = true
+}
+
+variable "allowed_locations" {
+  description = <<-EOT
+    Regions AI resources may live in (gcp.resourceLocations value groups or
+    region names, e.g. ["in:us-locations"] or ["us-east1", "us-central1"]).
+    Empty (the default) applies no residency constraint — residency is a policy
+    the operator states explicitly, not a default this module guesses.
+  EOT
+  type        = list(string)
+  default     = []
+}
+
+variable "enforce_require_cmek" {
+  description = <<-EOT
+    Require customer-managed encryption keys for the services below
+    (gcp.restrictNonCmekServices). Off by default: CMEK adds a KMS keyring and
+    rotation lifecycle that is pure operational surface at demo scale — the same
+    documented-and-deferred stance the factory takes on registry CMEK. Turn on
+    for regulated estates.
+  EOT
+  type        = bool
+  default     = false
+}
+
+variable "cmek_required_services" {
+  description = "Services that must use CMEK when enforce_require_cmek is on."
+  type        = list(string)
+  default = [
+    "aiplatform.googleapis.com",
+    "notebooks.googleapis.com",
+  ]
+}