OpenAI and Anthropic launch new safety tests for frontier AI models

OpenAI and Anthropic have introduced a new set of safety benchmarks designed to measure risks in the most capable AI models. The companies said the tests are meant to give researchers and policymakers a clearer way to evaluate whether advanced systems behave safely under pressure.

The benchmarks focus on frontier large language models, which are increasingly powerful and widely used across industries. As these systems become more capable, concerns have grown about harmful outputs, misuse, and other unintended behavior that could affect users and institutions.

The release adds to a broader debate over how best to assess AI safety before models are deployed at scale. Researchers have long argued that existing evaluation methods do not fully capture the ways advanced systems can fail in real-world settings.

By publishing the benchmarks together, OpenAI and Anthropic are signaling a push for more standardized testing in the fast-moving AI sector. The announcement is likely to draw attention from regulators, academic researchers, and companies racing to build and deploy more powerful models.