UK's AI Security Institute finds standard benchmarks systematically underestimate what AI agents can actually do

The UK's AI Security Institute found that standard AI evaluations underestimate agent capabilities by limiting the compute budget. Success rates in software engineering tasks increased by about 25% when the token budget was raised tenfold. Newer models benefit significantly from this increase. The findings suggest that actual progress at the frontier is about 60% steeper than previous measurements indicated, highlighting the need to reassess current evaluation standards. This reflects how compute budget constraints can skew our understanding of AI capabilities.