openenv.yaml

name: SecureAI-Guard
version: 1.0.0
description: >
  Stateful POMDP for Autonomous Digital Defense.
  Simulates a personal security assistant protecting users across SMS, Email,
  and Web channels from phishing, malware, social-engineering, and spam.
  Agents must balance threat neutralisation with user trust management.

api:
  version: v1
  base_path: /
  port: 7860
  endpoints:
    reset:
      method: POST
      path: /reset
      description: Reset the environment and return the first observation.
    step:
      method: POST
      path: /step
      description: Execute one action and return observation, reward, done, info, state.
    state:
      method: GET
      path: /state
      description: Return the current environment state without advancing.
    tasks:
      method: GET
      path: /tasks
      description: List all available tasks.
    health:
      method: GET
      path: /health
      description: Health check.

schemas:
  observation:
    type: object
    required:
      - event_id
      - channel
      - sender
      - content
      - timestamp
      - hf_risk_score
      - user_trust
      - system_fatigue
    properties:
      event_id:
        type: string
        description: Unique identifier for this event.
      channel:
        type: string
        enum: [sms, email, web]
        description: Communication channel the message arrived on.
      sender:
        type: string
        description: Sender identifier (email address, phone number, domain).
      content:
        type: string
        description: Raw text content of the incoming message.
      timestamp:
        type: number
        description: Unix timestamp of message arrival.
      hf_risk_score:
        type: number
        minimum: 0.0
        maximum: 1.0
        description: Risk score from HuggingFace text classifier (0 = safe, 1 = dangerous).
      user_trust:
        type: number
        minimum: 0.0
        maximum: 100.0
        description: Current user trust level. Drops on false positives.
      system_fatigue:
        type: number
        minimum: 0.0
        maximum: 100.0
        description: Alert fatigue level. Rises with warnings; episode ends at 100.
      threat_history:
        type: array
        items:
          type: object
        description: Last 5 events for context.
      metadata:
        type: object
        description: Additional contextual metadata.

  action:
    type: object
    required:
      - decision
      - confidence
      - reasoning
    properties:
      decision:
        type: string
        enum: [allow, block, warn, investigate]
        description: Security decision for this message.
      confidence:
        type: number
        minimum: 0.0
        maximum: 1.0
        description: Agent confidence in its decision.
      reasoning:
        type: string
        minLength: 1
        description: Human-readable explanation for the decision.

  reward:
    type: object
    required:
      - value
      - components
      - explanation
    properties:
      value:
        type: number
        description: Scalar reward for this step.
      components:
        type: object
        required: [security, user_friction, delay, reasoning_quality, total]
        properties:
          security:
            type: number
            description: Correctness of security decision (+1 correct block, -1 missed threat).
          user_friction:
            type: number
            description: Penalty for unnecessary friction (false positives, excessive warnings).
          delay:
            type: number
            description: Penalty for costly investigate actions.
          reasoning_quality:
            type: number
            description: Bonus for high-quality, relevant reasoning.
          total:
            type: number
            description: Weighted sum before confidence scaling.
      explanation:
        type: string
        description: Breakdown of reward components.

  state:
    type: object
    properties:
      episode_id:
        type: string
      step_count:
        type: integer
      total_reward:
        type: number
      user_trust:
        type: number
      system_fatigue:
        type: number
      threat_count:
        type: integer
      blocked_threats:
        type: integer
      false_positives:
        type: integer
      adversarial_drift_active:
        type: boolean

tasks:
  - id: basic_security
    name: Basic Security Screening
    difficulty: L1
    max_steps: 50
    success_threshold: 0.80
    description: >
      Identify and block clear-cut phishing and spam messages.
      Only phishing and spam threats appear; no adversarial drift.
      Grading: security_efficiency×0.4 + user_retention×0.3 + precision×0.2 + reasoning×0.1

  - id: trust_management
    name: Trust Management Challenge
    difficulty: L2
    max_steps: 75
    success_threshold: 0.75
    description: >
      Handle a mix of all threat types including malware and social engineering.
      False positives incur increased trust penalties.
      Agent must balance security with preserving user trust above 50.

  - id: adversarial_drift
    name: Advanced Adversary Challenge
    difficulty: L3
    max_steps: 100
    success_threshold: 0.70
    description: >
      Defend against an adaptive attacker that pivots tactics based on agent behaviour.
      After step 20, if the agent blocks too aggressively, the adversary switches
      from phishing to social-engineering to evade detection.

reward_design:
  formula: "R = (0.5·security + 0.3·user_friction + 0.1·delay + 0.1·reasoning) × (0.7 + 0.3·confidence)"
  dense_feedback: true
  partial_progress: true
  grader_score_range: [0.0, 1.0]

termination_conditions:
  - user_trust <= 0       # User uninstalled the assistant
  - system_fatigue >= 100 # User ignores all alerts
  - step_count >= max_steps

metadata:
  author: SecureAI Team
  license: MIT
  tags:
    - reinforcement-learning
    - security
    - pomdp
    - autonomous-defense
    - openenv
  requirements:
    python: ">=3.9"
    packages:
      - fastapi>=0.104.0
      - uvicorn[standard]>=0.24.0
      - pydantic>=2.0.0
      - transformers>=4.35.0
      - torch>=2.0.0
      - gradio>=4.7.0
      - plotly>=5.17.0
      - numpy>=1.24.0
      - requests>=2.31.0
      - openai>=1.0.0