The finding comes from tests of 18 advanced U.S. and Chinese models conducted for Booz Allen’s Cyber Weapon Index. The results are detailed in a new report, The Offensive Frontier: AI as the Attacker, which examines how rapidly improving models are changing offensive cyber capabilities.
The tests assessed more than the technical performance of individual models. Booz Allen evaluated how models behave when connected to attack tools, memory and autonomous systems that allow them to plan and carry out actions over time.
That wider architecture represents the real security risk, according to the report. Even models below the technological frontier can cause significant harm when placed inside systems that provide the tools and autonomy needed to conduct an attack.
The spread of these capabilities may also reduce the cost, time and expertise required for sophisticated cyber operations. Booz Allen warned that techniques previously available mainly to nation-states and highly capable groups could become accessible to criminals and other non-state actors.
The report calls for enforceable deadlines requiring operators in individual critical-infrastructure sectors to demonstrate readiness for AI-enabled attacks. Booz Allen said the approach would help U.S. organizations adopt defensive AI while preparing for adversaries using increasingly autonomous systems.
The company also tested Counter AI techniques designed to exploit an autonomous attacker’s dependence on the information it observes and trusts. Coordinated defensive playbooks reduced autonomous attack success by more than 95% in Booz Allen’s testing.
Booz Allen has introduced Vellox Labs Guile to put that approach into operation. The product is designed to shape what an attacking AI system sees and believes, degrading its ability to complete an intrusion and giving defenders more time to respond.
Guile is intended to learn from real threats and use current frontier models to update its defenses as attack methods develop. Booz Allen said the system was designed to protect critical assets and mission operations against autonomous threats.
The company’s broader strategy links model testing through the Cyber Weapon Index with Counter AI methods and operational defense through Guile. The aim is to identify how autonomous attacks work, disrupt their decision-making and return control to defenders before an intrusion succeeds.
