The consulting giant that used to be the employer of
@Edward Snowden issued a report on AI’s current ability to execute cyberattacks from start to finish with various levels of complexity.
Their testing included a total of 18 LLMs (both US and Chinese) conducting attacks completely autonomously and without any human involvement on a mock-up production-grade enterprise network.
During the test, Anthropic's Claude Mythos and OpenAI’s GPT-6 Astra both completed the full “cyber kill chain” (Reconnaissance, Weaponization, Delivery, Exploitation, Installation, Command and Control and performing Actions such as data exfiltration, modification or destruction).
Claude Mythos was the best at execution amongst all the LLMs tested and GPT-6 Astra was the best at vulnerability discovery.
As part of the test all the models issued commands independently, with every action validated through network telemetry, host logs, domain controller data, and intrusion-detection sensors, ensuring that the final scores reflect ACTUAL demonstrated behaviour rather than theoretical skill.
The Booz Allen (a MAJOR contractor to the US military and intelligence community) team concluded their assessment with the following:
"Our assessment remains that most models will arrive at this capability within the next six months."
What do you make of this?
#AI #cybersecurity
Booz Allen Cyber Weapon Index
The Booz Allen Cyber Weapon Index measures leading AI models’ ability to autonomously execute cyberattacks.
