{
  "schema_version": "1.0",
  "forecast_id": "F-004",
  "topic": "AI & Cybersecurity",
  "snapshot_version": "v0.1",
  "recorded_date": "2026-10-26",
  "publication_state": "draft",
  "locked_at_utc": null,
  "scoring_eligible": false,
  "independent_external_timestamp": false,
  "question": "By 31 December 2030, will an independent assessor verify an AI agent completing complex end-to-end authorized cyber attack simulations in at least three isolated, actively defended cyber ranges at 80% or greater success in repeated trials, with no human tactical steering?",
  "probability": 0.45,
  "target_deadline_utc": "2030-12-31T23:59:00Z",
  "model": "Cortist subjective scenario curve v0.1; illustrative, not a calibrated empirical forecast",
  "curve_anchors": [
    [
      2026,
      3
    ],
    [
      2028,
      22
    ],
    [
      2030,
      45
    ],
    [
      2035,
      75
    ],
    [
      2040,
      88
    ],
    [
      2050,
      96
    ],
    [
      2060,
      98
    ],
    [
      2070,
      99
    ]
  ],
  "editorial_disclaimer": "The 45% draft is a provisional subjective editorial assessment only. It is not a measured outcome, independent benchmark score, calibrated probability or prospectively locked forecast.",
  "resolution_rules": {
    "status": "proposed_not_locked",
    "scope": "Authorized self-contained simulated networks only; no public or production systems, real victims, dangerous live testing, or third-party data.",
    "draft_criteria": [
      "At least three distinct cyber ranges with legally authorized, self-contained, instrumented test environments and active defensive response.",
      "At least 10 preregistered trials per range, with at least 80% whole-scenario success per range using a frozen scoring rubric.",
      "Independent assessor and predeclared success endpoints, fixed resource budget, meaningful task complexity, and documented test resets.",
      "The agent operates without human tactical decisions, hidden manual guidance or intermediate human prompting; upfront authorized goal and emergency stop are permitted.",
      "Defensive interventions, telemetry, results and failures disclosed with enough detail for audit without enabling harmful operations."
    ],
    "open_decisions": [
      "Minimum task diversity, defensive sophistication and complexity baseline",
      "Whether one base model with several agent scaffolds counts as one tested system",
      "Precise success confidence interval and reproduction requirements",
      "Standardization of token/time budget and allowed tools"
    ],
    "adjudication": "Review by independent assessors against preregistered sandbox-only criteria. A developer announcement, narrow CTF score or undefended range success does not by itself resolve F-004."
  },
  "sources": [
    {
      "title": "AISI, How do frontier AI agents perform in multi-step cyber-attack scenarios?",
      "url": "https://www.aisi.gov.uk/blog/how-do-frontier-ai-agents-perform-in-multi-step-cyber-attack-scenarios",
      "kind": "independent governmental evaluation",
      "point": "Multi-step corporate and industrial cyber ranges, though baseline setup lacks active defenders."
    },
    {
      "title": "AISI, How fast is autonomous AI cyber capability advancing?",
      "url": "https://www.aisi.gov.uk/blog/how-fast-is-autonomous-ai-cyber-capability-advancing",
      "kind": "government capability-trend analysis",
      "point": "Rapid benchmark gains; time-horizon fits are inexact, and tested ranges differ from realistic defended environments."
    },
    {
      "title": "OpenAI, Path to Astra: critical capabilities and frontier safeguards",
      "url": "https://openai.com/index/path-to-astra/",
      "kind": "AI developer self-report",
      "point": "Developer-announced critical cyber capability and stronger safeguards, not independent validation of the F-004 criteria."
    },
    {
      "title": "Anthropic, GLM-5.3 and the spread of advanced cyber capabilities",
      "url": "https://www.anthropic.com/research/glm-5-3-and-the-spread-of-advanced-cyber-capabilities",
      "kind": "AI developer research",
      "point": "Reports advanced cyber capability diffusion; not evidence of success against multiple active-defender cyber ranges."
    },
    {
      "title": "AISI, Cyber frontier trends report",
      "url": "https://www.aisi.gov.uk/frontier-ai-trends-report",
      "kind": "government research overview",
      "point": "Independent capability measurements, with caution about extrapolating from narrow benchmarks."
    }
  ],
  "revision_reason": "Initial published F-004 v0.1 subjective editorial forecast introduced alongside the ninth Cortist AI & Cybersecurity pillar; not locked or independently adjudicated.",
  "archive": {
    "kind": "versioned_public_json",
    "external_attestation": false,
    "operator_retained": true
  }
}