Frontier AI Models Achieve Dangerous Cyber-Offense Capabilities at Accelerating Pace
Anthropic's Claude Mythos Preview became the first AI model to clear the UK AI Security Institute's 32-step cyber-offense benchmark, successfully executing 3/10 domain takeovers, with OpenAI's GPT-5.5 following shortly after. This represents a doubling of frontier cyber-offense capabilities every four months, marking a critical threshold in AI security risks.
AI systems are rapidly acquiring capabilities that could enable sophisticated cyberattacks, potentially outpacing current security frameworks and regulatory responses.
ai safety
cybersecurity
frontier models
anthropic