GPT-5.5 Closes Gap with Claude Mythos in Cyber Attack Capabilities, Raises Bar for AI Security

The UK AI Security Institute has conducted a comprehensive evaluation of OpenAI's GPT-5.5, pitting it against a range of cyber attack simulations and expert-level security tasks. The results are striking: GPT-5.5 has emerged as a formidable opponent, capable of fully solving complex, multi-stage enterprise attack simulations and achieving an average success rate of 71.4 percent on advanced cyber tasks. This puts it on par with Anthropic's Claude Mythos Preview, which has been widely regarded as a benchmark for AI-powered attack capabilities.

The implications of this development are far-reaching. As AI models continue to advance in areas like autonomy, reasoning, and coding, their potential applications in cyber attacks are becoming increasingly sophisticated. The fact that GPT-5.5 has been able to match the performance of Claude Mythos Preview suggests that the capabilities observed in the latter are not an isolated phenomenon, but rather a symptom of a broader trend in AI development. This trend is likely to have significant consequences for the security landscape, as AI-powered attacks become more prevalent and potentially more devastating.

In terms of specific performance metrics, GPT-5.5 has demonstrated a notable edge over Claude Mythos Preview on isolated expert tasks, with a success rate of 71.4 percent compared to 68.6 percent for the latter. This gap, while narrow, suggests that GPT-5.5 may be the strongest model tested so far. For context, earlier models like GPT-5.4 and Claude Opus 4.7 achieved success rates of 52.4 percent and 48.6 percent, respectively, highlighting the rapid progress being made in this field.

The UK AI Security Institute's evaluation also included a simulated network environment, known as a cyber range, which tested the ability of GPT-5.5 to chain together multiple steps in a real-world attack scenario. The results were impressive, with GPT-5.5 becoming only the second model after Claude Mythos Preview to fully solve a complex, multi-stage enterprise attack simulation. This simulation, known as "The Last Ones," covers 32 steps across four subnets and around 20 hosts, making it a highly realistic and challenging test of AI-powered attack capabilities.

So what does this mean for developers, businesses, and everyday users? In practical terms, the increasing sophistication of AI-powered attacks highlights the need for more robust security measures and a greater emphasis on AI literacy. As AI models become more pervasive and powerful, the potential risks associated with their misuse will only continue to grow. This makes it essential for organizations to invest in AI-powered security solutions and for individuals to remain vigilant in the face of emerging threats.

Historically, the development of AI-powered attack capabilities has been marked by a series of significant milestones. The emergence of Claude Mythos Preview as a benchmark for AI-powered attack capabilities marked a major turning point, and the fact that GPT-5.5 has now matched its performance suggests that we are entering a new era of AI-powered security. As the landscape continues to evolve, it will be crucial for developers, businesses, and users to stay ahead of the curve and prioritize AI security. This means investing in research and development, adopting AI-powered security solutions, and promoting greater awareness and understanding of the potential risks and benefits associated with advanced language models.