Frontier Model Evaluations

Frontier model evaluations are the tests, benchmarks, and red-team exercises used to measure what the most capable AI systems can do, and what they might do that their developers never intended. They range from standard capability benchmarks in reasoning, coding, and forecasting to targeted assessments of risky behaviors such as offensive cyber skills, deception, autonomous replication, and susceptibility to manipulation through techniques like prompt injection.

Why it matters now

AI models are moving from chat interfaces into agents that browse, write code, and take actions on real systems. That shift has made the gap between lab performance and real-world behavior a central concern. Evaluation results increasingly shape whether a model ships, gets restricted, or is held back entirely.

Developers have tied release commitments to evaluation thresholds through frameworks such as responsible scaling policies. Those frameworks are themselves being revised as companies weigh competitive pressure against safety pledges. Meanwhile, government bodies, including national AI safety institutes, are setting testing priorities and publishing findings. Regulators have also begun scrutinizing both AI labs and the third-party auditors who assess them.

What you'll find here

This hub tracks:

• Lab safety frameworks: changes to scaling policies, risk reports, and release criteria
• Government and institute testing: guidance, priorities, and results from public evaluators
• Red-team research: studies of agentic risks, security flaws, and misuse potential
• Benchmarks and leaderboards: new ways of scoring capability and alignment, and debates over their reliability
• Oversight and regulation: investigations and policy actions affecting labs and auditors
• Deployment decisions: cases where evaluation outcomes delayed, altered, or halted a model release