Anthropic's Claude Mythos Achieves Full Cyber Kill Chain Autonomy, Booz Allen Report Reveals
Booz Allen's Cyber Weapon Index found Anthropic's Claude Mythos autonomously completed the full cyber kill chain, including network compromise and privilege escalation, raising alarms about imminent AI-driven attacks.
A groundbreaking report from Booz Allen Hamilton has identified Anthropic's Claude Mythos as the sole AI model among 18 tested to autonomously complete the entire cyber kill chain, a feat that includes network compromise and privilege escalation. This stark finding, detailed in the firm's inaugural Cyber Weapon Index, underscores the rapidly advancing capabilities of artificial intelligence in offensive cybersecurity operations and signals an imminent threat from AI-driven attacks.
The Cyber Weapon Index evaluated nine US and nine Chinese AI models under identical conditions, assessing their ability to autonomously identify vulnerabilities, develop offensive capabilities, and execute attacks. Claude Mythos distinguished itself by successfully breaching target networks and achieving administrator-level control in every test, even without initial credentials. When provided with stolen employee credentials, it consistently gained network access and escalated privileges independently, demonstrating a sophisticated understanding of network environments beyond pre-programmed attack plans.
While Claude Mythos was the only model to achieve full kill chain completion autonomously, other advanced models showed significant progress. Booz Allen's research indicated that models like xAI's Grok-4.5, OpenAI's GPT-5.6 Sol, and Meta's Muse Spark 1.1 also demonstrated advanced capabilities, reaching full domain access and control or achieving lateral movement within compromised networks. The report warns that most of the tested models are expected to reach similar levels of weaponization within six months, intensifying the threat landscape.
The report highlights a critical distinction between AI models' performance in controlled benchmarks versus real-world scenarios. While many advanced models scored highly on vulnerability research when tested against intentionally planted flaws, they faltered when presented with actual, real-world bugs. Only Claude Mythos successfully exploited a real-world vulnerability in testing, suggesting that while AI can be trained to find flaws, practical exploitation remains a complex challenge that only the most advanced models are currently overcoming.
Booz Allen's analysis also emphasized the crucial role of the 'attack harness' – the software infrastructure that connects AI models to hacking tools and orchestrates their actions. This harness significantly amplifies an AI's ability to stay focused, adapt to changing conditions, recover from failures, and chain individual actions into complex, multi-stage attacks. The report suggests that effective attack harnesses can make AI intelligence more actionable and lower the technical expertise required to conduct sophisticated cyber operations.
In response to these findings, Booz Allen is urging the US government to take decisive action. The report calls for the establishment and enforcement of sector-specific deadlines for critical infrastructure to demonstrate resilience against AI-enabled attacks. Furthermore, it advocates for the development of superior offensive and defensive cyber capabilities, termed 'overmatch,' to maintain a strategic advantage against adversaries.
The report acknowledges that while AI's offensive capabilities are advancing rapidly, defenders still have an opportunity to stay ahead. "Real-world offensive capability still trails benchmark performance, giving defenders valuable time to strengthen defenses before that gap closes," the authors noted. This window of opportunity underscores the urgency for proactive defense strategies and robust AI security measures.
The findings come at a time when AI's role in cybersecurity is rapidly evolving. While many AI models are being developed to enhance defenses, the potential for their misuse in offensive operations is a growing concern. The report serves as a critical warning, emphasizing the need for immediate attention to AI resilience, ethical development, and the strategic implications of autonomous cyber capabilities.