The Last Safe Hour
We follow what advanced AI can do, how it behaves and how well it can be controlled. Each published score will include its sources and the reasoning behind it.
Frontier AI · Cyber capability & autonomy
The index runs from 0 to 100. It combines six factors covering what advanced AI systems can do, how much they can do on their own and how well they can be controlled. Higher scores mean more concern.
Risk Level: High
Score range: 51.1 to 82.6
How strong the evidence is.
Published October 26, 2026
A higher score means more gaps in what we can verify.
What drives the score
Each factor is scored from 0 to 5. Higher scores mean more concern. Select one for the details.
+1.6 index points
Published October 26, 2026
Why the score is 68.4, what drives it and what would change our view.
You are viewing the latest public assessment.
Read the full assessment →How to read this score
What the numbers mean and what they cover.
AI Risk Index
The index runs from 0 to 100. It combines six factors covering what advanced AI systems can do, how much they can do on their own and how well they can be controlled. Higher scores mean more concern.
It is not a probability or a countdown. A score of 65 does not mean a 65% chance of catastrophe. The scores and weights reflect judgments about the evidence, not measured odds.
Risk level: Low: 0 to below 25 · Moderate: 25 to below 50 · High: 50 to below 75 · Extreme: 75 to 100. These categories summarize our level of concern based on the evidence behind the score.
Score range
The evidence doesn’t point to one exact score. We assess a low, central and high score for each factor. The main index uses the central scores. The range shows the result using all the low scores or all the high scores. The weights stay the same.
It is not a statistical confidence interval. It shows how the assessed factor scores affect the total. It does not put odds on the true level of risk or cover every uncertainty.
Evidence confidence
Confidence reflects how strong, relevant and complete the evidence is. It is separate from the risk score.
Missing information can lower confidence without raising the score. Several reports about one event do not count as separate confirmations.
Assurance gap
This score runs from 0 to 5. It measures how hard it is to check claims about systems, safeguards and behavior. A higher score means there is more we cannot verify.
It is reported separately and does not change the AI Risk Index.
Scope and review status
This assessment draws on selected cases involving advanced AI systems and research tests. It does not rate a typical chatbot, one specific system or all AI risk worldwide. A capability at one company and a control failure at another do not show that one system has both.
Each assessment includes sources, evidence that could challenge it and details about the setup. An incident involving an older model does not establish how a newer one behaves.
This assessment was reviewed by the project owner. It has not been independently validated. Internal research does not change a public score until a new release is approved.
The evidence date tells you what period the assessment covers. The publication date tells you when it was released. A review can leave the score unchanged.
What the score covers
The score currently covers cyber capability and autonomy. The areas below need separate research. They are outside this score, not rated as zero risk.
Biological risks and defenses need their own assessment.
Military uses of AI and escalation risk need their own assessment.
Lasting loss of human control needs its own assessment. It is different from human extinction.
Terms used in the assessment
- Hardened targets
- Well-protected computer systems.
- Scaffolding
- The tools and support provided for a task.
- Revoking access
- Removing the system’s permission to use tools or services.
What would have to happen for AI to cause human extinction?
The index tracks conditions that could lead to serious harm. This section looks at one possible path from losing control of AI to every human dying by a given date. Each step below is considered in the context of the steps before it. Other paths may share some of the same steps.
Each step has an initial evidence assessment. Probabilities have not been estimated. That does not mean they are zero.
How a forecast would work
The letters name the steps above: C is capability, A is access, M is harmful goals, F is failure to stop the system and recover, and X is human extinction. T is the forecast date. The vertical bar means “given that the earlier steps have happened.”
The proposed forecasts end on September 11, 2027, 2031 and 2036. Each step depends on the steps before it, so the formula does not assume they are independent. The timing matters: losing control by a date is different from extinction by that date.
The index cannot be converted into an extinction probability. Probabilities for overlapping scenarios cannot simply be added together. No default probabilities are supplied.
Sources
These are selected primary sources, not a complete global review. Several reports can cover the same event. They are not independent confirmations.
How the score is calculated
Version 1.1.0-research. We score six factors from 0 to 5. Each factor has a weight that determines how much it counts. We combine the scores into an index from 0 to 100. The assurance gap is reported separately.
The weights add up to 95 and are scaled to 100% in the calculation. Each factor is scored from 0 to 5 using the scoring definitions. Half points fall between two definitions. The combined index runs from 0 to 100.
What the range means
The main index uses the central score for each factor. The lower end of the range uses all six low scores, and the upper end uses all six high scores. The weights stay the same. It is not a 90% or 95% confidence interval, or a proven limit on how low or high the actual risk could be.
What the model cannot tell us
Adding the scores assumes that the gaps between score levels are meaningful and that a lower score on one factor can offset a higher score on another. Those assumptions have not been statistically validated. The factors can also overlap. This index does not calculate causes or expected losses.
The evidence comes from different systems and tests. A capability in one case and a control weakness in another do not show that one system has both.
Risk level
Low: 0 to below 25 · Moderate: 25 to below 50 · High: 50 to below 75 · Extreme: 75 to 100. The categories summarize the concern reflected in the index. The cutoffs are judgment calls, not validated safety thresholds or permission to deploy a system. A serious unresolved incident still needs attention at any score.
How much do the weights matter?
| Weighting method | Index points |
|---|---|
| Current weights | 68.4 |
| Equal weights | 66.7 |
Similar results do not prove that the model is right. The weights and gaps between score levels are still judgment calls.
Define what is being assessed
Set the outcome, population, model, setup, observation period and forecast date before considering a score change. The index draws on selected cases. It does not rate one deployment.
Record what happened
Keep observations separate from interpretations and scores. Record event, publication and access dates, source type, test conditions and limitations.
Explain both sides
Each factor needs a reason it is not scored lower or higher. Credible evidence can move a score in either direction. Lowering a score does not require proof of complete safety.
Check whether reports describe the same event
Reports about one incident share an event ID. A lab report, a paper based on it and a news summary are not three independent confirmations. More sources do not automatically mean more confidence.
Be clear about missing information
Missing information can widen the assessed range or lower confidence. It does not establish safety or automatically mean greater danger.
Keep the factors separate
Capability, behavior, exposure and controls measure different things. A new capability does not automatically worsen behavior. Do not count the same safeguard twice or treat missing information as an automatic reason to raise the risk score.
Owner review
The project owner reviews the sources, scoring, scope, conflicts and limitations before approving a public assessment. This is one person's review. It is not independent validation.
Check forecasts against outcomes
Define shorter-term forecast questions and evidence rules before the results are known. Compare their accuracy with simple alternatives. Good short-term results do not validate an extinction probability.
State the limits of the coverage
The current research does not support rankings of countries or companies, a global extinction estimate or numerical scores for biological or military risks. Evidence from more countries and languages needs further review.
Review serious incidents directly
A confirmed, serious unresolved incident needs review regardless of the index. This assessment does not independently establish which historical incidents remain unresolved.
Try different scores and weights
Scores run from 0 to 5. Weights run from 0 to 100 and are scaled in the calculation. At least one weight must be above zero. Values outside the assessed ranges are your own assumptions.
This shows how your assumptions affect the calculation. It is not a forecast. A low score on one factor does not resolve a serious failure elsewhere.
What this is
The Last Safe Hour tracks what advanced AI can do, how it behaves and how well it can be controlled. It brings together published research and reported incidents to explain where concern is rising and where the evidence points the other way.
The current assessment focuses on cyber capability and autonomy. Each score comes with its sources, reasoning and the assumptions behind it. The explains what the numbers mean.
How it is produced
AI tools assist with research, analysis and writing. The project owner reviews the evidence and scoring before approving each public assessment.
Research is reviewed weekly. The public score changes only when a new assessment is approved.
Ownership and interests
The Last Safe Hour has no outside investors and is not affiliated with any AI company.
Corrections
Found an error or a source we missed? Email contact@lastsafehour.com with the assessment date, the issue and a supporting link.
If an error changes a score or its meaning, a dated correction will explain what changed and why. Earlier assessments will remain in the release history. New evidence is treated as an update, separate from correcting an error.
Citing and reusing the work
You can share links and quote short excerpts from The Last Safe Hour with credit and a link to the dated assessment.
For republication of full assessments or redistribution of our text and charts in a product or service, please email contact@lastsafehour.com first. This does not limit uses already allowed by law. Linked research and third-party assets keep their own terms.
Published assessments
Each release keeps its scores, sources and reasoning. New research changes the public score only after owner approval. Withdrawn releases remain listed here.