Researchers say AI just broke every benchmark for autonomous cyber capability - CyberScoop

Researchers from the UK's AI Security Institute and Palo Alto Networks have found that Anthropic's Claude Mythos Preview and OpenAI's GPT-5.5 have substantially exceeded the historical doubling trend in autonomous cybersecurity task completion. The findings were published on Wednesday in a joint report, which analyzed the models' performance on a standard set of cybersecurity benchmarks designed to measure autonomous offensive and defensive capabilities.

According to the report, the new models scored significantly higher than previous generations, suggesting a capability leap that outpaces the roughly six-month doubling cycle observed in earlier models. The benchmarks included tasks such as vulnerability discovery, exploit chaining, and achieving defined objectives in simulated network environments. The researchers noted that the improvement was 'substantially beyond' what the trend would predict.

The implications of the findings are significant for both offensive and defensive cybersecurity. The UK AI Security Institute stated that the results underscore the need for enhanced monitoring and guardrails around autonomous AI systems. Palo Alto Networks said the study demonstrates the importance of proactive defense. Neither organization called for a pause in development, but both recommended further research into the models' capabilities and limitations.