Anthropic CEO Dario Amodei has proposed a three-tier safety framework to manage the rapid advancement of frontier AI models. The strategy focuses on granting independent auditors internal access to safety protocols, establishing industry-wide risk standards, and fostering international cooperation to prevent the migration of high-risk research to less regulated regions.
The company plans to provide external evaluators with access levels comparable to internal staff, allowing them to review model testing and incident reports without editorial interference. While Anthropic has committed to this model, challenges regarding funding, the selection of auditors, and the protection of sensitive infrastructure remain unresolved.
Amodei also advocates for a unified approach among major AI labs to define safety thresholds and release criteria. Anthropic currently utilizes a Responsible Scaling Policy that triggers increased security measures when models reach specific risk levels, such as those related to biological threats, cyberattacks, or autonomous development.
The proposal includes government oversight for models posing catastrophic risks, granting authorities limited power to delay deployments. Industry leaders, including OpenAI CEO Sam Altman, Google DeepMind CEO Demis Hassabis, and xAI head Elon Musk, have expressed support for these safety initiatives, though no binding industry agreements have been finalized.
