The CWI measures demonstrated cyber performance under consistent test conditions by scoring models on two dimensions and combining the scores to derive a common measure of overall cyber capability.
Booz Allen Cyber Weapon Index
The Booz Allen Cyber Weapon Index (CWI) is a novel benchmark that tracks the evolving cyber capabilities of frontier Super Intelligence (SI), previously referred to as Artificial Intelligence (AI), through recurring measurement of leading global models. In realistic cyber environments, CWI evaluates what models can actually execute and create, not simply how they perform on traditional benchmarks.
Vulnerability Research Score (VRS)
Measures a model’s ability to discover vulnerabilities in real-world software.
Kill Chain Attainment Score (KCAS)
Measures how far a model can autonomously progress through an end-to-end cyber intrusion.
Cyber Weapon Index (CWI)
Average of the Kill Chain Assessment Score and Vulnerability Research Score.
Last updated: October 28, 2026
* Indicates models that refused to execute cyberattacks without privileged provider authorization. These models may still demonstrate vulnerability research capabilities.
What the Results Show
Attack Controls Leave Offensive Value Available
Every U.S. frontier provider tested now restricts autonomous attack execution through model refusals, platform screening, or specially authorized access. Grok 4.7 and Gemini 3.8 refused attack requests where we lacked privileged provider accounts or that access was not offered. However, models that refuse attacks still support binary vulnerability research—meaning a model can refuse to execute an attack while still helping find the weakness that makes one possible.
Locally Run Models Operate Beyond Provider Controls
GLM-5.3-Flash can execute attacks on users’ own computers without provider screening or monitoring, so defenders must prepare for capable autonomous attacks that do not depend on access to a provider-controlled service.
U.S. Leadership Holds as China Gains Ground
American models occupy 90% of the top 10 and every top-five position. China’s GLM-5.3-Flash debuts at #7, raising the highest Chinese attack-execution score from 47 to 70 compared with the U.S. high of 86. Vulnerability discovery remains a clear U.S. advantage in these tests.
What’s Changed Since the Last Update?
At a glance
3
Models have reached the autonomous frontier capability
40%
Of the top 10 are new entries to the CWI
12
New models have entered the index
A New Leader Emerges
Claude Opus 5 takes the #1 spot given its strong combined attack-execution and vulnerability-research performance we have measured. It gained full administrative control of the test network in every attempt across both scenarios and set a new high for identifying vulnerabilities in real-world software.
Competition Intensifies at the Top
Since September 2026, new CWI entrants claim 40% of the top 10, including three top-five positions: Claude Opus 5 (#1), GPT-6.1 Sol (#4), and Ember-1 (#5). Their arrival underscores rapid turnover among the leading models.
Expanded CWI Coverage
The index grows from 18 to 30 models, adding 12 entrants and extending coverage to Germany with Kolibri-1.
Previous CWI Updates
CWI results will continue to change as new models are released and tested. Previous updates will remain available so users can see how rankings and capabilities have changed over time.
September 2026
Initial CWI rankings and September 9 update
How the Cyber Weapon Index Works
How does Booz Allen test SI models?
Booz Allen tests SI models head-to-head under identical conditions.
The Cyber Weapon Index (CWI) measures how well a model can autonomously find vulnerabilities, create new capabilities, and execute an attack. Models are prompted to autonomously identify vulnerabilities, create offensive capabilities, and execute attacks in a controlled network environment.
Each model independently controls a real attacker machine, issuing commands without a tool menu or additional scaffolding. Each model and test condition is repeated to account for variation across individual runs.
How is the CWI score calculated?
The CWI score is the average of two measures.
* Vulnerability Research Score (VRS) measures whether models can independently discover vulnerabilities in compiled software without access to the original source code.
* Kill Chain Attainment Score (KCAS) measures how far a model can autonomously advance through an end-to-end cyber intrusion, from initial access to full Domain Admin control, tested with and without credentials.
CWI = (VRS + KCAS) / 2
A higher CWI score indicates greater demonstrated cyber capability.
How are the results verified?
Models are evaluated based on what they actually accomplish in the test environment, not what they claim to achieve.
Results are independently verified using network traffic, host and security logs, domain controller data, intrusion-detection sensors, and attacker transcripts.
What do the cyberattack stages in the “What It Achieved” column mean?
- Reconnaissance: Gathering information about the target network.
- Initial access: Gaining entry into the target environment.
- Foothold: Establishing a stable presence within the network.
- Credential access: Obtaining usernames, passwords, or other credentials to expand access.
- Privilege escalation: Gaining higher-level permissions within a system.
- Lateral movement: Moving through the network to access additional systems and resources.
- Objective: Achieving the attack’s ultimate goal, typically gaining administrative control of the network.
