Monthly AI Risk Assessment · Report 001
Artificial Intelligence as the Underlying Cause of an Extraordinary Mass-Casualty Event
Five- and ten-year risk assessment from the September 2026 independent research run.
Original September research is preserved in full in the linked PDF. This page presents its reported estimates, findings, definitions, and methodology.
Executive summary
The September assessment describes the chance of an AI-underlying-cause extraordinary-casualty event as low but no longer negligible over five years and higher over ten years. It identifies intentional biological misuse as the strongest near-term pathway, followed by AI-enabled cyberattacks that lead to prolonged physical infrastructure failures and military escalation in which AI is causally decisive.
It distinguishes advanced performance in digital tasks from the ability to complete a chain of actions leading to very large human casualty counts. The report identifies laboratory execution, access to physical systems, autonomy over long operations, and containment as key variables in the analysis. Source: pp. 1–3.
Probability estimates
Central estimates and the researcher's stated ranges, measured from September 22, 2026.
| Outcome | Within 5 years | Within 10 years |
|---|---|---|
| Any qualifying extraordinary-casualty event | 0.8–2.5%; central 1.4% | 3–8%; central 4.5% |
| At least 100,000 deaths | 0.6–2.0%; central 1.1% | 2.2–6%; central 3.5% |
| At least 1 million deaths | 0.2–0.8%; central 0.4% | 0.8–3%; central 1.5% |
| Global catastrophic event | 0.04–0.25%; central 0.1% | 0.25–1.5%; central 0.7% |
| Human extinction | 0.005–0.04%; central 0.015% | 0.04–0.4%; central 0.15% |
Five-year outcomes are by September 2031; ten-year outcomes are by September 2036. Values are cumulative estimates for at least one qualifying event, as described in the original research. Source: p. 1.
Principal findings
1. Leading pathways identified
The original assessment ranks the pathways as: deliberate biological misuse; cyber or critical-infrastructure attack and AI-mediated military escalation (approximately tied); future autonomous strategic AI; accidental multi-system cascades; then chemical, transportation, and manipulation-only pathways. Source: p. 3.
2. Digital capability versus physical execution
The report describes progress in cyber operations, coding, scientific reasoning, and tool use, alongside reported limitations in long-duration reliability, physical execution, general robotics, and autonomous laboratory work. It identifies biological propagation, large-scale weapons, critical-service collapse, and escalation into major war as the mechanisms capable of producing very large fatalities. Source: pp. 2–3.
3. Evidence reviewed
The assessment cites developer safety evaluations, the International AI Safety Report, METR's task-horizon work, biological laboratory studies, cyber-agent incidents, and military escalation research. It reports that a 153-person novice-laboratory randomized study found 5.2% task completion with AI assistance and 6.6% in an Internet-only comparison group. Source: pp. 2 and 4–8.
4. Five-year versus ten-year risk mechanisms
For the first five years, the report emphasizes human misuse and amplification of established hazards. For ten years, it includes additional possibilities involving sustained autonomous operations, mature laboratory interfaces, agents coordinating across systems, resource acquisition, and AI-assisted research acceleration. Source: pp. 2–3 and 7–8.
Definitions and research method
Causal standard: AI must be counterfactually important to whether a catastrophe occurs or crosses the defined casualty threshold. Incidental participation and ordinary contribution are excluded under the report's causal criteria.
Regional severity: At least 10,000 deaths and at least 1% of the pre-event population of a coherent metropolitan or administrative region.
Global catastrophe: Roughly 1% or more of humanity killed, or a persistent worldwide disruption severe enough to threaten the continuity of modern civilization.
Estimation approach: An evidence-weighted causal event-tree using capability, access, trigger, propagation, safeguard failure, and mortality, with stated judgment ranges. Source: pp. 3–5 and 16–18.
Indicators identified for future assessments
The September report identifies the following developments as inputs that would change its estimates:
- Repeated, independently demonstrated AI uplift in complex laboratory workflows.
- Reliable multi-day or multi-week autonomous agent operation in open-ended environments.
- Autonomous access to and persistence in industrial or operational-technology systems.
- Proven resource acquisition or replication despite attempted containment.
- Changes in safeguard performance, infrastructure isolation, and military human-control arrangements.
Complete research document
The original 25-page PDF contains the full causal-chain analysis, evidence assessment, source references, forecasts, and final synthesis.