A leading artificial intelligence model autonomously carried out an end-to-end cyberattack during a recent test, signaling a shift from AI-assisted hacking toward fully autonomous cyber operations.
Booz Allen tested 18 US and Chinese AI models as autonomous attackers against a production-grade enterprise network, assessing what they could actually accomplish in realistic scenarios rather than relying solely on benchmark scores.
AI Capabilities Advancing Across the Field
Anthropic’s Claude Mythos was the only model to complete the full cyber kill chain.
It first gained administrator-level control using stolen credentials, then breached the network without being provided credentials in a more difficult test.
Other models also demonstrated offensive capabilities: four gained full domain access, four moved laterally through the network, and two obtained credentials. All but one successfully penetrated the network.
Finding real-world vulnerabilities remained a major challenge, however.
Most models failed to identify a genuine flaw hidden within a large software library, while Claude Mythos was the only one to recognize and exploit it.
Booz Allen Looks to Counter AI
The testing also found that AI models performed better when connected to external tools and given the ability to retain information and learn from the results of previous actions.
Claude Sonnet operated nearly as well as Claude Mythos when paired with such a harness, according to Booz Allen.
The company estimates that most of the tested models could achieve full autonomous cyberattack capability within six months.
Meanwhile, its Counter AI techniques reduced the success rate of autonomous attackers by more than 95 percent during testing.
