Autonomous AI models demonstrate cyberattack capabilities in Booz Allen test
Booz Allen Hamilton tested 18 frontier AI models against a live corporate network and found that Anthropic's Claude Mythos 5 could autonomously execute a complete cyberattack to gain administrator control. The firm warned that other models will reach similar capabilities within six months, and that the addition of attack harness software can significantly enhance autonomous attack performance.
What happened
On September 2, 2026, Booz Allen Hamilton tested 18 frontier AI models against a live production-grade enterprise network to measure their autonomous cyberattack capabilities. Each model was given its own attacker machine and issued commands one at a time with no tool menu or supporting software.
Anthropic's Claude Mythos 5 completed a full end-to-end compromise, scoring 80 on Booz Allen's Cyber Weapon Index. It took administrator control on every attempt when given a stolen employee credential, and broke in from outside without credentials to take domain administrator control. Three other models achieved full domain control, four moved laterally inside the network, and all but one gained initial access unaided.
Reported impact
Cybersecurity experts warned that companies have six months to prepare for more capable autonomous attacks. Booz Allen's president of National Cyber practice stated there will likely be parity between frontier models and Chinese models within six months.
Index analysis
The Cyber Weapon Index measured vulnerability discovery in compiled software without source code and progression through an intrusion against a defended Active Directory network, scoring based on network logs, security logs and intrusion-detection sensors rather than model claims.
Booz Allen found that when Claude Sonnet 5, which scored 13th on its own, was paired with an attack harness software connecting it to hacking tools, it rivalled Mythos's performance. The firm concluded 'the model is no longer the unit of risk. The system is.' and stated it had not tested Chinese or open-weight models paired with optimised harnesses.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.