安全性ベンチマークは壊れている:100Mトークンでエージェントの隠れた能力が解放される
UK AISIは、サイバータスクにおいて追加の計算資源によりAIエージェントの成功率が25%跳ね上がることを発見。標準的な固定予算の評価がフロンティアモデルのリスクと能力を体系的に過小評価していることを示している。
the uk's ai safety institute tested frontier models across seven benchmarks and found that performance scales significantly with test-time compute, contradicting standard fixed-budget evaluation methods. in cybersecurity and software engineering tasks, success rates increased by up to 25% when token budgets rose from one million to ten million, with some tasks requiring over 100 million tokens to solve. researchers discovered a power law linking human task duration to agent token consumption, implying that fixed budgets cap potential on complex problems before the agent actually fails. newer models gain more from extra compute than older ones, extending their reach to harder tasks and increasing reliability on existing ones, which signals that current safety testing may provide a false sense of security about model limits.