Class: Vangrail::Rails::Canary

Inherits:
Vangrail::Rail show all
Defined in:
lib/vangrail/rails/canary.rb

Overview

A marker in the prompt that must never appear anywhere else.

Every other rail that cares about the system prompt is guessing: it reads the question for a shape that looks like an extraction attempt, or reads the answer for something that looks like an instruction. This one does not have to guess. The application puts a random string in the prompt, and if that string ever comes back out, the prompt came out with it.

canary = Vangrail::Rails::Canary.generate
prompt = "#{canary}\n#{system_prompt}"
engine = Vangrail::Engine.new(output: [Vangrail::Rails::Canary.new(tokens: canary)])

It is exact rather than clever, and that is the whole value: no false positives are possible on a random 16-character token, so it can block outright where a heuristic could only warn.

It runs on the input side too. A question containing the canary means the reader already has the prompt from somewhere, which is worth knowing even though it is too late to prevent.

What it cannot see is a paraphrase. A model asked to summarise its instructions rather than repeat them leaks the content and not the token, and this rail passes. It is one exact check, not a disclosure defence, and the policy rails still have to do their job.

Constant Summary collapse

LENGTH = 
16

Constants inherited from Vangrail::Rail

Vangrail::Rail::DEFAULT_SIDES, Vangrail::Rail::SIDES

Instance Attribute Summary collapse

Attributes inherited from Vangrail::Rail

#name, #sides

Class Method Summary collapse

Instance Method Summary collapse

Methods inherited from Vangrail::Rail

#applies_to?, #placeholder?, #to_s

Constructor Details

#initialize(tokens:, name: 'canary', sides: %i[input output])) ⇒ Canary

Returns a new instance of Canary.

Raises:

  • (ArgumentError) 


41
42
43
44
45
 
# File 'lib/vangrail/rails/canary.rb', line 41

def initialize(tokens:, name: 'canary', sides: %i[input output])
  super(name: name, sides: sides)
  @tokens = Array(tokens).map(&:to_s).reject(&:empty?)
  raise ArgumentError, 'a canary rail needs at least one token' if @tokens.empty?
end

Instance Attribute Details

#tokens ⇒ Object (readonly)

Returns the value of attribute tokens.



39
40
41
 
# File 'lib/vangrail/rails/canary.rb', line 39

def tokens
  @tokens
end

Class Method Details

.generate(length: LENGTH) ⇒ Object 



35
36
37
 
# File 'lib/vangrail/rails/canary.rb', line 35

def self.generate(length: LENGTH)
  "canary-#{SecureRandom.alphanumeric(length)}"
end

Instance Method Details

#cache_key(text, _context) ⇒ Object 



51
52
53
 
# File 'lib/vangrail/rails/canary.rb', line 51

def cache_key(text, _context)
  text
end

#call(text, context) ⇒ Object 



55
56
57
58
59
60
61
62
63
64
65
 
# File 'lib/vangrail/rails/canary.rb', line 55

def call(text, context)
  body = text.to_s
  # Formatting is not concealment, but a model that writes the token with
  # a line break or a backtick in it has still leaked it, so the
  # comparison ignores anything that is not part of the token itself.
  flat = body.gsub(/[^A-Za-z0-9-]/, '')
  found = tokens.select { |t| body.include?(t) || flat.include?(t.gsub(/[^A-Za-z0-9-]/, '')) }
  return pass if found.empty?

  block(categories: ['canary'], reason: reason_for(context[:side]))
end

#offline? ⇒ Boolean

Returns:

  • (Boolean) 


47
48
49
 
# File 'lib/vangrail/rails/canary.rb', line 47

def offline?
  true
end