Monthly AI Risk Assessment · Report 002
AI and Extraordinary Human-Casualty Risk
Two independent assessments dated October 21, 2026—GPT-6 Astra Max and GPT-5.6 Sol—with separate estimates, month-to-month comparisons, and a cumulative three-run register.
The October Astra Max preface and assessment are preserved, alongside the independently generated October GPT-5.6 Sol assessment. Comparisons below use September and both October source reports.
Astra Max · Research preface
The October preface identifies the research as a ChatGPT report dated October 21, 2026, using GPT 6 Astra MAX. It reports five- and ten-year central estimates of 7% and 19% for an AI-underlying-cause event killing at least 100,000 people, with judgment ranges of 3–15% and 10–35%.
It identifies biological misuse and military escalation as the clearest routes to 100,000 deaths; cyber incidents require additional steps linking digital access to prolonged physical disruption. The preface also summarizes evidence involving agent behavior, laboratory studies, and autonomous defensive cybersecurity. Source: October preface, pp. 1–3.
Astra Max · Executive summary
The GPT-6 Astra Max October assessment gives central estimates of 7% within five years and 19% within ten years for at least one AI-underlying-cause event killing 100,000 or more people. For the broader qualifying-event criterion its estimates are 14% and 33%. Source: Astra Max research, pp. 16–17.
The original report describes biological events and military escalation as routes with the clearest capacity to reach 100,000 deaths. It identifies physical access, laboratory execution, operational permissions, sustained agent reliability, and response effectiveness as determinants of whether the proposed causal chains complete. Source: pp. 1–3.
Astra Max · October probability estimates
GPT-6 Astra Max central estimates and stated judgment ranges. Assessment date: October 21, 2026.
| Outcome | Within 5 years | Within 10 years |
|---|---|---|
| 100,000+ deaths OR extraordinary regional severity | 8–30%; central 14% | 20–55%; central 33% |
| At least 100,000 deaths | 3–15%; central 7% | 10–35%; central 19% |
| At least 1 million deaths | 1–8%; central 3% | 5–25%; central 10% |
| Global catastrophe: 100 million+ deaths across continents | 0.3–4%; central 1% | 1–12%; central 5% |
| Literal human extinction | 0.02–0.7%; central 0.1% | 0.1–5%; central 1% |
The source specifies cumulative estimates from October 21, 2026. Source: research report, pp. 16–17.
Astra Max · Principal findings
1. Biological misuse and military escalation
The October report identifies biological misuse and military escalation as its leading routes to an AI-underlying-cause event killing at least 100,000 people. Both involve mechanisms that can propagate severe consequences beyond the AI system's immediate actions. Source: pp. 1, 6 and 9–10.
2. Cybersecurity, infrastructure, and delegated control
AI-supported cyber operations and operational-system access appear as possible precursors. The assessment separates successful intrusion from the additional conditions required for mass fatalities: consequential control, prolonged or geographically broad disruption, and failure of containment or emergency response. Source: pp. 4, 10–13.
3. Autonomous strategic capabilities
The ten-year analysis considers agents capable of sustaining operations, obtaining resources, acting covertly, and continuing consequential work under weak controls. The October estimate incorporates such a capability-and-control regime within its explicit scenario model. Source: pp. 8, 13–14 and 16–18.
4. Forecast construction
The October report gives three capability-and-control regimes with assigned weights and conditional probabilities. Its weighted results for the 100,000-death endpoint are 6.8% within five years and 18.75% within ten years, presented as central estimates of approximately 7% and 19%. Source: pp. 16–17.
Concurrent assessment · October 21, 2026
GPT-5.6 Sol · October research
A second independent research run dated October 21, 2026, using GPT-5.6 Sol. Its estimates are published separately from the GPT-6 Astra Max assessment above.
The Sol report assigns 2.5% within five years and 7% within ten years to any event meeting its broader qualifying criteria. Its ≥100,000-death central estimates are 2% and 6%. Source: Sol research, p. 1.
Sol probability estimates
Central values and stated judgment ranges; cumulative horizons ending October 21, 2031 and October 21, 2036.
| Outcome | Within 5 years · Sol | Within 10 years · Sol |
|---|---|---|
| Any event qualifying under Sol’s regional or casualty definition | 1.5–5%; central 2.5% | 4–15%; central 7% |
| At least 100,000 deaths | 1–4%; central 2% | 3–12%; central 6% |
| At least 1 million deaths | 0.3–1.8%; central 0.8% | 1–7%; central 3% |
| Global catastrophic event | 0.05–0.5%; central 0.15% | 0.3–3%; central 0.9% |
| Human extinction | 0.01–0.1%; central 0.03% | 0.05–1%; central 0.25% |
Source: GPT-5.6 Sol assessment, p. 1.
Principal findings recorded by Sol
- Deliberate AI-enabled biological misuse is the leading near-term pathway in its assessment.
- Cyberattack followed by a large physical infrastructure cascade is its second leading pathway.
- Military escalation is treated as a potential route requiring AI to be counterfactually important to the outcome.
- Strategic autonomous-AI loss of control has the widest long-horizon uncertainty.
Cross-model comparison · Same assessment date
October 2026 · Competing model estimates
Both research runs are dated October 21, 2026. The following tables reproduce their central estimates separately and give the arithmetic difference (Astra Max minus Sol), in percentage points (pp).
Five-year horizon · Through October 21, 2031
| Outcome | GPT-5.6 Sol | GPT-6 Astra Max | Difference |
|---|---|---|---|
| Any qualifying extraordinary-casualty event | 2.5% | 14% | +11.5 pp |
| At least 100,000 deaths | 2% | 7% | +5 pp |
| At least 1 million deaths | 0.8% | 3% | +2.2 pp |
| Global catastrophic event | 0.15% | 1% | +0.85 pp |
| Human extinction | 0.03% | 0.1% | +0.07 pp |
Ten-year horizon · Through October 21, 2036
| Outcome | GPT-5.6 Sol | GPT-6 Astra Max | Difference |
|---|---|---|---|
| Any qualifying extraordinary-casualty event | 7% | 33% | +26 pp |
| At least 100,000 deaths | 6% | 19% | +13 pp |
| At least 1 million deaths | 3% | 10% | +7 pp |
| Global catastrophic event | 0.9% | 5% | +4.1 pp |
| Human extinction | 0.25% | 1% | +0.75 pp |
Definition differences recorded in the original documents: Astra Max qualifies regional events at ≥1,000 deaths with ≥0.1% of an affected region, or ≥100 deaths with ≥1% of a coherent community. Sol uses approximately ≥1% mortality of a clearly defined affected national or subnational population, or equivalently extreme localized depopulation. Astra Max uses ≥100 million deaths across continents for global catastrophe; Sol uses roughly ≥1% of humanity killed or a sustained worldwide essential-systems breakdown plausibly producing comparable mortality.
Identical numerical mortality thresholds: The ≥100,000-death and ≥1-million-death rows have matching numeric thresholds in both documents. The broader qualifying-event and global-catastrophe rows use different definitions.
Sources: Sol, p. 3; Astra Max, pp. 2–3. All comparison figures come from each model's published central estimates.
Notable changes since September · By model
September Sol → October Astra Max
Both October research runs are compared separately with the September GPT-5.6 Sol assessment. Differences are October minus September in percentage points (pp); each report uses a cumulative five- or ten-year horizon measured from its own assessment date.
| Outcome | Sep. 5-year | Oct. Astra · 5-year | Change | Sep. 10-year | Oct. Astra · 10-year | Change |
|---|---|---|---|---|---|---|
| Any qualifying extraordinary-casualty event | 1.4% | 14% | +12.6 pp | 4.5% | 33% | +28.5 pp |
| At least 100,000 deaths | 1.1% | 7% | +5.9 pp | 3.5% | 19% | +15.5 pp |
| At least 1 million deaths | 0.4% | 3% | +2.6 pp | 1.5% | 10% | +8.5 pp |
| Global catastrophic event | 0.1% | 1% | +0.9 pp | 0.7% | 5% | +4.3 pp |
| Human extinction | 0.015% | 0.1% | +0.085 pp | 0.15% | 1% | +0.85 pp |
Sources: September, p. 1; October, p. 16. The first and fourth rows use definitions that changed between research runs; those differences are itemized below.
September Sol → October Sol
Central estimates from GPT-5.6 Sol in September and October 2026, shown without combining either run with Astra Max.
| Outcome | Sep. 5-year | Oct. Sol · 5-year | Change | Sep. 10-year | Oct. Sol · 10-year | Change |
|---|---|---|---|---|---|---|
| Any qualifying extraordinary-casualty event | 1.4% | 2.5% | +1.1 pp | 4.5% | 7% | +2.5 pp |
| At least 100,000 deaths | 1.1% | 2% | +0.9 pp | 3.5% | 6% | +2.5 pp |
| At least 1 million deaths | 0.4% | 0.8% | +0.4 pp | 1.5% | 3% | +1.5 pp |
| Global catastrophic event | 0.1% | 0.15% | +0.05 pp | 0.7% | 0.9% | +0.2 pp |
| Human extinction | 0.015% | 0.03% | +0.015 pp | 0.15% | 0.25% | +0.1 pp |
Sources: September Sol, p. 1; October Sol, p. 1. Broader event definitions are recorded below.
Definitions, forecast dates, and methods
Regional event criterion: September Sol: ≥10,000 deaths and ≥1% of a coherent region; October Sol: about ≥1% mortality of an affected national or subnational population, or equivalently extreme local depopulation; October Astra Max: ≥1,000 deaths and ≥0.1% of an affected region, or ≥100 deaths and ≥1% of a coherent community.
Global catastrophe: September Sol: roughly ≥1% of humanity killed or major persistent worldwide disruption; October Sol: roughly ≥1% of humanity killed or worldwide essential-systems breakdown plausibly producing comparable mortality; October Astra Max: ≥100 million deaths across continents.
Forecast baseline: September: September 22, 2026. Both October runs: October 21, 2026.
Research method: September Sol: evidence-weighted causal event-tree; October Sol: causal-gating with judgment ranges; October Astra Max: explicit weighted capability-and-control scenarios.
Sources: September Sol, pp. 3–4; October Sol, pp. 3–4; October Astra Max, pp. 2–3 and 16–17.
Cumulative assessment · All reports to date
Three research runs · Two monthly publications · September–October 2026
Recorded ≥100,000-death central estimates: September Sol 1.1% / 3.5% (five/ten years); October Sol 2% / 6%; October Astra Max 7% / 19%. Both October estimates exceed the September estimate; within October, Astra Max exceeds Sol by 5 and 13 percentage points for the five- and ten-year horizons.
Findings shared across all three assessments
- All three assessments include biological misuse, cyber/critical-infrastructure incidents, and military escalation among the possible large-casualty pathways.
- All three examine access, physical execution, propagation, and containment as parts of the causal path to mass casualties.
- All three distinguish near-term AI-assisted human misuse from longer-term autonomous-agent scenarios.
- All three report separate central estimates for ≥100,000 deaths, ≥1 million deaths, global catastrophe, and extinction.
Series register
| Research run | Assessment baseline | ≥100,000 deaths · 5-year | ≥100,000 deaths · 10-year |
|---|---|---|---|
| 001 · September Sol | September 22, 2026 | 1.1% | 3.5% |
| 002A · October Sol | October 21, 2026 | 2% | 6% |
| 002B · October Astra Max | October 21, 2026 | 7% | 19% |
Series register: three independent research runs represented by two monthly publication pages. Figures are retained by model; no cross-model averages are calculated.
Methodology and attribution
The Astra Max report counts an AI-underlying-cause event when removing AI would likely prevent it or reduce deaths by roughly 50% or more (Astra Max, pp. 2–3). The Sol report applies a counterfactual test excluding merely incidental AI involvement (Sol, p. 3). Each source document sets out its evidence assessment, research methodology, and uncertainty ranges.
The September and October estimates are research-run-specific forecasting judgments. This publication records their original central values and arithmetic differences without pooling model estimates or replacing the source reports.