Frontier AI Models Cross Cyber-Offense Threshold, Capability Doubling Accelerates
Anthropic's Claude Mythos Preview became the first AI model to successfully complete complex cyber-attack simulations, achieving partial domain takeovers, with OpenAI's GPT-5.5 matching this capability within weeks. The UK AI Security Institute now estimates cyber-offense capabilities are doubling every four months, accelerating from seven months in late 2025.
This represents a critical inflection point where AI systems transition from theoretical cyber threats to demonstrated offensive capabilities, fundamentally altering cybersecurity risk calculations.
cybersecurity
frontier ai
ai safety
cyber offense