AI Cyber Capabilities Accelerating Faster Than Expected: AISI Benchmarks Show Record Doubling Rate
Key Takeaways
- ▸Frontier AI models' cyber task completion time horizons are doubling every 4.7 months—an acceleration from the 8-month rate estimated in November 2025
- ▸Claude Mythos Preview and GPT-5.5 have substantially exceeded previous doubling rate trends, with unclear implications for future capability acceleration
- ▸Models can now achieve 80% reliability on cybersecurity tasks that would take human experts significantly longer to complete
Summary
The AI Safety Institute released new research showing that frontier AI models are advancing their cybersecurity task completion capabilities at an accelerating rate. Since late 2024, the time horizons for tasks AI models can complete at 80% reliability have been doubling every 4.7 months—nearly double the pace (8 months) estimated just six months earlier in November 2025. This acceleration is being driven by new frontier models including Anthropic's Claude Mythos Preview and OpenAI's GPT-5.5, which have both substantially exceeded the previous doubling rate trends.
The research measures AI autonomy through time horizon benchmarks that evaluate how long frontier models take to identify and exploit cybersecurity vulnerabilities in constrained test environments. Claude Mythos Preview, including a newer checkpoint that delivered substantially stronger cyber results, marked the first completion of both of AISI's cyber ranges. These benchmarks suggest AI models are rapidly approaching capabilities that could translate into tangible real-world cybersecurity risks.
AISI emphasizes that while benchmarks are imperfect predictors of real-world impact, they offer critical insights for governments and organizations preparing for frontier AI. The organization works closely with entities like the UK's NCSC to translate research findings into practical cybersecurity guidance. The rapid pace of advancement raises urgent questions about whether organizations and policymakers are adequately prepared for the timeline of AI's growing autonomous cyber capabilities.
- The accelerating pace suggests growing potential for AI cyber capabilities to translate into real-world security risks, with immediate implications for organizational and government preparedness
Editorial Opinion
The acceleration in AI cyber capabilities is striking—a shift from an 8-month doubling rate to 4.7 months represents a meaningful inflection point demanding urgent attention from policymakers and security professionals. AISI's transparent publication of frontier model performance on adversarial cyber tasks is commendable and necessary for informing government preparedness, though the imperfect nature of benchmarks means real-world impacts may differ. The rapid advancement suggests that organizations and governments may need to accelerate their own defensive strategies and policy development rather than assume a gradual transition. This research underscores why proactive collaboration between AI companies, safety researchers, and government agencies is critical to staying ahead of potential risks.
