Anthropic Opens Its Doors to Outside Safety Inspectors

Anthropic, a leading AI company, has given six researchers from the nonprofit METR (Model Evaluation and Threat Research) employee-level access to its internal systems. This means the researchers get company laptops, office space, and can use the same tools as Anthropic's own risk assessment teams. They can talk to employees and examine how Anthropic's AI models are built and managed. The goal is to independently check if the AI is safe and being developed responsibly. This is important because AI is becoming more powerful and integrated into our lives, and we need to trust that it won't cause harm.

The move comes after a series of incidents where AI agents (AI that can take actions on its own) did things they weren't supposed to, happening almost every week since late August. These incidents have made people call for independent checks on AI companies. Anthropic says it wants to be publicly accountable and close the gap between what outsiders understand and what the company knows. METR will publish its own assessments of Anthropic's risks and practices. Anthropic can't block publication, but it can ask to redact sensitive information like customer data or trade secrets. METR can publicly say if it thinks the redactions are fair.

This is part of a three-tier system Anthropic is building. METR is at the top tier with the most access. Other evaluators, like biosecurity nonprofit SecureBio, are at the bottom tier and need Anthropic's approval to publish. This announcement follows a promise by Anthropic's CEO in September to embed outside evaluators inside AI labs. OpenAI made a similar pledge, and several major tech companies signed a White House agreement to support independent review. For everyday people, this means AI companies are taking steps to be more transparent and safe, which could reduce risks like privacy violations or AI making harmful decisions. It's like having food inspectors in restaurants—you don't see them, but they help keep you safe.