A study by the UK's AI Security Institute shows that common benchmarks fail to accurately assess AI agents' capabilities, particularly under limited computing budgets. The findings indicate that performance improves significantly with increased compute time, especially in cybersecurity and software development tasks.