# Rails Security

> Ruby on Rails security practices for Rails 7.1+ and 8.x apps. Covers production configuration, authentication, authorization, strong parameters, SQL injection, XSS and output escaping, CSRF, file uploads, API security, credentials, untrusted boundaries, and dependency scanning. Use when reviewing or writing Rails authentication, authorization, input handling, uploads, or deployment configuration.

- Skill: `affaan-m/rails-security` (Agent Skill)
- Install (CLI): `npx get-skill add affaan-m/rails-security`
- Raw SKILL.md: https://api.skillmd.com/api/skills/affaan-m/rails-security/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Data & Analytics
- Author: Affaan Mustafa (https://skillmd.com/u/affaan-m)
- Updated: 2026-11-18
- Page: https://skillmd.com/skills/affaan-m/rails-security

---


# Rails Security

Security practices for Ruby on Rails applications (Rails 7.1+ and 8.x). Rails is secure by
default in most places; the vulnerabilities that reach production usually come from opting
out of a default rather than failing to add protection. This skill is the "how." For the
decisions about which control to reach for, see the Ruby security rules —
`rules/ruby/security.md` in this repository, installed as `rules/ecc/ruby/security.md`.

## When to Activate

- Writing or reviewing authentication, sessions, or account recovery
- Adding authorization checks to controllers or views
- Handling user input, file uploads, or webhook payloads
- Configuring a Rails app for production
- Reviewing a PR that touches `config/`, `app/controllers/`, or `Gemfile`
- Any use of `html_safe`, `raw`, `Arel.sql`, or raw SQL strings

## Production Configuration

```ruby
# config/environments/production.rb
Rails.application.configure do
  config.force_ssl = true
  # Only use when all requests pass through a proxy that rejects or redirects HTTP.
  config.assume_ssl = true # behind a proxy that terminates TLS

  # Reject requests with unexpected Host headers.
  config.hosts << "example.com"
  config.hosts << /.*\.example\.com/

  config.action_dispatch.cookies_same_site_protection = :lax
  config.log_level = :info
end
```

Filter sensitive values out of logs:

```ruby
# config/initializers/filter_parameter_logging.rb
Rails.application.config.filter_parameters += %i[
  passw secret token _key crypt salt certificate otp ssn cvv card_number
]
```

## Authentication

Rails 8 ships an authentication generator that produces session-based auth with password
reset and metadata tracking:

```bash
bin/rails generate authentication
```

Use Devise when you need OAuth, MFA, confirmable, lockable, or multi-model auth. For
hand-rolled auth, `has_secure_password` handles bcrypt hashing:

```ruby
class User < ApplicationRecord
  has_secure_password

  validates :email, presence: true, uniqueness: { case_sensitive: false }
  validates :password, length: { minimum: 12 }, allow_nil: true
end
```

`authenticate_by` looks up and verifies in one call, so a missing user and a wrong password
take the same time:

```ruby
# Bad: the timing difference reveals whether the email exists
user = User.find_by(email: params[:email])
if user&.authenticate(params[:password])

# Good
user = User.authenticate_by(email: params[:email], password: params[:password])
```

Rotate the session on sign-in and on any privilege change:

```ruby
class SessionsController < ApplicationController
  def create
    user = User.authenticate_by(email: params[:email], password: params[:password])
    return render :new, status: :unprocessable_entity unless user

    reset_session
    session[:user_id] = user.id
    redirect_to root_path
  end

  def destroy
    reset_session
    redirect_to root_path
  end
end
```

### Account recovery

Password reset tokens must expire, be single-use, and not reveal whether an account exists.
Signed tokens generated by the model give you both:

```ruby
class User < ApplicationRecord
  generates_token_for :password_reset, expires_in: 15.minutes do
    password_salt&.last(10) # invalidates the token once the password changes
  end
end

user = User.find_by_token_for(:password_reset, params[:token])
```

Render the same response whether or not the address matched.

## Authorization

Authentication is not authorization. Every action that touches a record must check that the
current user may act on it.

```ruby
# Bad: any signed-in user can load any invoice
@invoice = Invoice.find(params[:id])

# Good: scope through the owner
@invoice = current_user.invoices.find(params[:id])
```

For anything past ownership, use a policy object (Pundit, Action Policy, or CanCanCan):

```ruby
class InvoicePolicy < ApplicationPolicy
  def update?
    user.admin? || record.user_id == user.id
  end
end

class InvoicesController < ApplicationController
  def update
    @invoice = Invoice.find(params[:id])
    authorize @invoice
    # ...
  end
end
```

Fail closed — verify that authorization ran rather than trusting each action to remember:

```ruby
class ApplicationController < ActionController::Base
  include Pundit::Authorization

  after_action :verify_authorized, except: :index
  after_action :verify_policy_scoped, only: :index
end

# Public actions have no record to authorize, so exempt them explicitly.
class SessionsController < ApplicationController
  skip_after_action :verify_authorized
end
```

## Strong Parameters and Mass Assignment

Never pass raw `params` to a model. Permit an explicit list:

```ruby
# Bad: a crafted request can set admin
User.new(params[:user])

# Good
def user_params
  params.require(:user).permit(:email, :name)
end
```

Rails 8 adds `expect`, which raises on a malformed structure instead of returning an empty
hash:

```ruby
def user_params
  params.expect(user: [:email, :name])
end
```

Attributes such as `admin`, `role`, `user_id`, and `account_id` should never appear in a
permit list driven by request data. Assign them explicitly from the session or a checked
policy.

## SQL Injection

Active Record parameterizes by default. Injection appears when a string is built by hand:

```ruby
# Bad
User.where("email = '#{params[:email]}'")
Order.order("#{params[:sort]} #{params[:dir]}")

# Good
User.where(email: params[:email])
User.where("email = ?", params[:email])
```

`order`, `pluck`, `group`, and `select` accept raw SQL and are not sanitized. Allowlist the
input rather than escaping it:

```ruby
SORTABLE = { "created_at" => "created_at", "total" => "total" }.freeze
DIRECTIONS = { "asc" => "ASC", "desc" => "DESC" }.freeze

column = SORTABLE.fetch(params[:sort], "created_at")
direction = DIRECTIONS.fetch(params[:dir], "DESC")
Order.order(Arel.sql("#{column} #{direction}"))
```

`LIKE` needs its own escaping, since `%` and `_` are wildcards:

```ruby
term = User.sanitize_sql_like(params[:q])
User.where("name LIKE ?", "%#{term}%")
```

## XSS and Output Escaping

ERB escapes interpolated values by default. Every bypass is security-sensitive:

```erb
<%# Bad: renders attacker-controlled markup %>
<%= @comment.body.html_safe %>
<%= raw @comment.body %>

<%# Good: escaped %>
<%= @comment.body %>

<%# Good: allowlisted markup when the field is genuinely rich text %>
<%= sanitize @comment.body, tags: %w[p br strong em a], attributes: %w[href] %>
```

`html_safe` does not sanitize — it marks a string as already safe and skips escaping. Only
call it on strings you constructed.

Interpolating into JavaScript needs the matching escape, not the HTML one:

```erb
<script>
  const payload = <%= raw @data.to_json %>;              // Bad
  const payload = <%= raw json_escape(@data.to_json) %>; // Good
</script>
```

`link_to`'s href accepts `javascript:` URLs, so validate user-supplied links:

```ruby
uri = URI.parse(url) rescue nil
safe = uri.is_a?(URI::HTTP) # URI::HTTPS subclasses URI::HTTP
```

Set a Content Security Policy:

```ruby
# config/initializers/content_security_policy.rb
Rails.application.configure do
  config.content_security_policy do |policy|
    policy.default_src :self
    policy.script_src  :self
    policy.object_src  :none
    policy.frame_ancestors :none
  end
end
```

## CSRF

Rails enables CSRF protection for new apps. Keep it on, and do not exempt actions to make a
client work:

```ruby
# Bad: disables protection for every action in the controller
skip_before_action :verify_authenticity_token
```

If an endpoint genuinely cannot carry the token — a third-party webhook, say — authenticate
it another way (signature verification) and scope the exemption to that action alone.

API-only controllers inheriting from `ActionController::API` have no cookie session, so CSRF
does not apply; token or session-cookie APIs still need it.

## File Uploads

The `content_type` a client sends is attacker-controlled, and `blob.content_type` is not a
substitute for it. Active Storage derives that column through `Marcel::MimeType.for`, which
treats the client's declared type as one candidate alongside magic-byte detection and keeps
it when the bytes match nothing known — so opaque bytes declared `image/png` are stored as
`image/png` with `identified?` true. Inspect the stored bytes yourself:

```ruby
module Uploads
  Rejected = Class.new(StandardError)

  MAGIC = {
    "application/pdf" => "%PDF-".b,
    "image/png" => "\x89PNG\r\n\x1a\n".b,
    "image/jpeg" => "\xFF\xD8\xFF".b
  }.freeze
  ALLOWED_EXTENSIONS = %w[.pdf .png .jpg .jpeg].freeze
  MAX_BYTES = 10.megabytes

  def self.verified!(blob)
    # Fail closed: an object that is missing or still pending cannot be inspected.
    raise Rejected, "could not be verified" unless blob.service.exist?(blob.key)
    raise Rejected, "is too large" if blob.byte_size > MAX_BYTES

    extension = File.extname(blob.filename.to_s).downcase
    raise Rejected, "extension not allowed" unless ALLOWED_EXTENSIONS.include?(extension)

    head = blob.service.download_chunk(blob.key, 0...8)
    raise Rejected, "type not allowed" unless MAGIC.any? { |_mime, magic| head.start_with?(magic) }

    blob
  end
end
```

The check reads from storage, so it has to run where the bytes already exist. A form upload
is not yet in storage when model validations run — Active Storage uploads on save — so
upload explicitly first, and attach only what passed:

```ruby
# Form upload
blob = ActiveStorage::Blob.create_and_upload!(io: params[:file], filename: params[:file].original_filename)
document.file.attach(Uploads.verified!(blob))

# Direct upload: the client already stored the object, so verify before the record
# points at it.
document.file.attach(Uploads.verified!(ActiveStorage::Blob.find_signed!(params[:file])))
```

Attach to a private service — a public bucket makes every upload world-readable regardless
of the checks above.

```ruby
has_one_attached :file, service: :private_documents
```

`rails_blob_url` and `url_for(attachment)` are stable references that never expire; they
redirect to a short-lived service URL, so anyone holding one keeps access. Serve
user-scoped files from your own controller that authorizes each request, with
`config.active_storage.draw_routes = false`, which removes every default Active Storage
route; apps using direct uploads need a replacement route:

```ruby
get "documents/:id/file", to: "document_files#show", as: :document_file
# Rails' direct-upload endpoint has no authentication of its own. This example uses a
# hand-rolled session sign-in; see Devise or the Rails 8 generator for their equivalents.
constraints ->(req) { req.session[:user_id].present? } do
  post "uploads", to: "active_storage/direct_uploads#create", as: :rails_direct_uploads
end

class DocumentPolicy < ApplicationPolicy
  def show?
    record.user_id == user.id
  end
end

class DocumentFilesController < ApplicationController
  def show
    document = current_user.documents.find(params[:id])
    authorize document, :show?
    redirect_to document.file.blob.url(expires_in: 5.minutes), allow_other_host: true
  end
end
```

Never interpolate a user-supplied path into `send_file`:

```ruby
# Bad: path traversal
send_file Rails.root.join("storage", params[:name])
```

## API Security

Rate-limit authentication and other abusable endpoints. Action Controller ships
`rate_limit` from Rails 7.2; use `rack-attack` on 7.1 or for app-wide throttling:

```ruby
class SessionsController < ApplicationController
  rate_limit to: 10, within: 3.minutes, only: :create
end
```

Verify webhook signatures with a constant-time comparison:

```ruby
class WebhooksController < ApplicationController
  skip_before_action :verify_authenticity_token, only: :create

  def create
    expected = OpenSSL::HMAC.hexdigest("SHA256", ENV.fetch("WEBHOOK_SECRET"), request.raw_post)
    given = request.headers["X-Signature"].to_s

    unless ActiveSupport::SecurityUtils.secure_compare(expected, given)
      return head :unauthorized
    end

    event_id = JSON.parse(request.raw_post).fetch("id")
    delivery = WebhookDelivery.create_or_find_by!(provider: "billing", delivery_id: event_id)
    return head :ok if delivery.processed_at?

    ProcessWebhookJob.perform_later(delivery.id, request.raw_post)
    head :ok
  end
end
```

```ruby
add_index :webhook_deliveries, %i[provider delivery_id], unique: true
```

The job marks `processed_at` when it finishes. Jobs retry, so `perform` must be
idempotent for the same reason. Requires a stable id in the signed body.

Keep CORS narrow — never reflect an arbitrary origin alongside credentials:

```ruby
# Bad
origins "*"

# Good
origins "https://app.example.com"
```

Serialize responses explicitly so new columns do not leak:

```ruby
# Bad: ships every column, including ones added later
render json: @user

# Good
render json: @user.as_json(only: %i[id name created_at])
```

## Secrets and Credentials

Secrets belong in Rails credentials, environment variables, or a secret manager:

```bash
bin/rails credentials:edit --environment production
```

`config/credentials/production.key` must stay out of version control. Read secrets through
`Rails.application.credentials` or `ENV.fetch` — `ENV[...]` returns `nil` and fails silently
when a variable is missing:

```ruby
# Bad
ENV["STRIPE_SECRET"]

# Good
Rails.application.credentials.dig(:stripe, :secret)
ENV.fetch("STRIPE_SECRET")
```

If a key is ever committed, rotate it. Removing the commit does not un-expose it.

## Untrusted Boundaries

Controller params are not the only entry point. Job arguments, webhook payloads, Action
Cable messages, and Turbo Stream targets all carry outside data:

- Pass record IDs to jobs and re-scope inside `perform`; a queued job runs without the
  original request's session.
- Authorize in `ApplicationCable::Connection` and in each channel's `subscribed`.
- Never build a Turbo Stream target from user input.

```ruby
class ExportJob < ApplicationJob
  def perform(user_id, invoice_id)
    user = User.find(user_id)
    invoice = user.invoices.find(invoice_id) # scoped, not Invoice.find
    # ...
  end
end
```

## Dependency Scanning

Run these when the lockfile changes, and in CI:

```bash
bundle exec bundle-audit check --update
bundle exec brakeman --no-progress
```

Brakeman flags the patterns above statically. Treat new warnings as blocking, and record
deliberate exceptions in `config/brakeman.ignore` with a reason rather than lowering the
confidence threshold.

## Verification Checklist

- [ ] `force_ssl` and `config.hosts` set in production
- [ ] Sensitive parameters filtered from logs
- [ ] Sessions rotated on sign-in and privilege change
- [ ] Every record lookup scoped to the current user or checked by a policy
- [ ] No permit list contains a role, ownership, or account column
- [ ] No raw SQL string interpolation; `order` and `pluck` inputs allowlisted
- [ ] Every `html_safe` and `raw` reviewed
- [ ] CSRF exemptions scoped to specific actions with another auth mechanism
- [ ] Uploads validated against stored bytes, not `blob.content_type`, and fail closed when the object is unavailable
- [ ] Auth endpoints rate-limited; webhook signatures compared in constant time
- [ ] No secrets in version control; master key excluded
- [ ] `bundle-audit` and `brakeman` clean

## Related Skills

- `rails-patterns` — the architectural patterns these controls sit inside
- `security-review` — secure-by-default review patterns across languages
- Ruby security rules (`rules/ruby/security.md`, installed as `rules/ecc/ruby/security.md`)
  — the decisions and when-to-use guidance this skill implements

