openenv.yaml

name: soc-incident-response
version: "1.0.0"
description: >
  A Security Operations Center (SOC) environment where an AI agent acts as a
  tier-1 analyst. The agent triages SIEM alerts, reconstructs attack chains
  using the MITRE ATT&CK framework, isolates compromised endpoints, and
  manages real business constraints during active incidents.
author: "Ankit Kumar & Jayanth"
tags:
  - openenv
  - cybersecurity
  - soc
  - incident-response
  - mitre-attack
  - real-world

tasks:
  - id: alert_triage
    name: "Alert triage (easy)"
    description: >
      Given a single SIEM alert with indicators of compromise, classify it as
      true positive or false positive and recommend the correct immediate action.
    difficulty: easy
    max_steps: 10
    score_range: [0.0, 1.0]

  - id: attack_chain_reconstruction
    name: "Attack chain reconstruction (medium)"
    description: >
      Correlate 15 alerts across 4 hosts over a 2-hour window. Identify the
      full MITRE ATT&CK kill chain, locate patient zero, and contain the threat
      before it reaches the crown jewel asset.
    difficulty: medium
    max_steps: 25
    score_range: [0.0, 1.0]

  - id: constrained_incident_response
    name: "Constrained incident response (hard)"
    description: >
      Respond to an active breach with business constraints: CEO laptop involved
      cannot isolate, domain controller at risk, legal hold required, and a
      customer-facing API is degrading. Balance security, business continuity,
      and compliance simultaneously.
    difficulty: hard
    max_steps: 40
    score_range: [0.0, 1.0]

observation_space:
  type: object
  description: >
    Structured snapshot of the SOC environment at each step. Contains the
    current alert queue, network topology state, active sessions, investigation
    history, and any business constraint flags.

action_space:
  type: object
  description: >
    Seven analyst actions: enrich_alert, correlate_alerts, isolate_endpoint,
    disable_account, collect_forensics, escalate_to_tier2, create_ticket.

reward:
  type: float
  range: [-1.0, 1.0]
  description: >
    Dense reward signal. Positive for correct classifications, chain coverage,
    and successful containment. Negative for false positive isolations,
    missed true positives, and per-minute dwell time penalties.

endpoints:
  reset: POST /reset
  step: POST /step
  state: GET /state