AI models are outrunning safety tests, leaving regulators with less visibility into their capabilities

AI safety and risk are increasingly moving beyond what can be measured. Frontier models from OpenAI, Anthropic and other labs have begun to get around the tests designed to assess their offensive cyber capabilities. This creates a critical gap in our understanding of what these models can do and what risks they carry.

The benchmark problem

Benchmarks for evaluating the cybersecurity of AI models were developed only a few years ago and rely on the following approaches:

• Screen capture and attempts to break into web applications
• Identifying vulnerabilities in security systems
• Social engineering and phishing scenarios

But frontier models now solve these tasks with a success rate that exceeds the accuracy of the tests themselves.

A regulatory vacuum

• US federal agencies must prepare a classified registry by August 1
• Current tools are insufficient for measuring real capabilities
• Risk: regulators may make decisions based on incorrect information

What this means

This evaluation gap poses a unique challenge for AI regulation. The tests that are supposed to measure safety are ceasing to work. That may mean the real risks from advanced AI systems are significantly higher than regulators understand. New evaluation methods are needed that keep pace with the growth of AI capabilities.