Researchers at the UK’s AI Security Institute found that standard benchmarks don’t accurately reflect the capabilities of AI agents, especially when computing budgets are limited. As a result, current evaluations may not show the full potential of these systems. When given more computing time, AI models’ success rates increased significantly, with notable gains in cybersecurity and software development tasks. This has significant implications for how we assess AI performance. Why it matters: this study highlights the need for more nuanced evaluation methods to accurately measure AI capabilities, which is crucial for advancing AI research and development.