Booz Allen report flags AI models capable of autonomous end-to-end cyber intrusions

Booz Allen published a Cyber Weapon Index report warning that frontier AI can now run fully autonomous, end-to-end network intrusions.

Tests of 18 U.S. and Chinese models found one could gain initial access and then take full network control without human guidance.

The report says the cost and expertise needed for advanced attacks have dropped sharply, widening access beyond nation-states to criminal groups.

The analysis shifts the risk focus from the model itself to the full system, citing tool use, memory, attack harnesses and autonomy as key amplifiers.

Counter-AI playbooks cut autonomous attacker success by more than 95% in testing; the product launch targets disrupting what attackers see and trust.