UK AI Security Institute Flags Major Benchmark Breach as Claude and GPT Models Surge Past Cybersecurity Projections

Two of the most advanced artificial intelligence systems have demonstrated substantial progress in autonomous cybersecurity operations, according to independent assessments published by the United Kingdom's AI Security Institute and Palo Alto Networks. Anthropic's Claude Mythos Preview and OpenAI's GPT-5.5 have significantly exceeded the doubling trajectory that researchers had been tracking since late 2024, raising questions about whether the pace of advancement in AI cyber capabilities is accelerating beyond earlier forecasts.

The AISI, which conducts pre-deployment evaluations of frontier models on behalf of the British government, estimated earlier this year that frontier models' 80% reliability cyber time horizon—a benchmark measuring how long a task takes a human expert, used as a proxy for AI autonomy—had been doubling approximately every five months. That rate itself represented roughly half the doubling period observed in prior cycles. Both the Claude Mythos Preview and GPT-5.5 have now surpassed those trend lines, though officials caution that it remains unclear whether the results reflect a singular capability leap or signal the beginning of a faster developmental trajectory.

The findings place added pressure on governments and security agencies to reassess existing frameworks for evaluating AI systems before deployment. Palo Alto Networks' parallel research corroborates the AISI conclusions, suggesting the capability gains are not confined to a single evaluation environment. Security researchers have long debated when AI systems might reliably match or exceed human performance across the full spectrum of cybersecurity tasks. The latest results intensify that debate without resolving it, leaving policymakers to weigh the implications of accelerating autonomy against the need for robust oversight mechanisms.