Anthropic's New Benchmarks Show Advanced AI Models Dramatically Improving at Developing Software Exploits

Anthropic researchers tested AI models on three new academic benchmarks designed to measure their ability to develop software exploits, finding that Mythos Preview significantly outperformed competing models. The findings suggest that as AI capabilities advance, the technical barrier to creating exploits will lower substantially, potentially democratizing a capability currently requiring specialized expertise.

Why it matters: Security teams and policymakers need to understand how rapidly AI is closing the gap on exploit development—a capability that could reshape cybersecurity risk if advanced model access spreads widely.