Volvo broadens Nvidia partnership to upgrade automated driving
The first vehicle using the new technology is due within the decade, but Volvo has not named the model or disclosed the partnership's value.
Anthropic gives METR employee-like access for AI safety reviews
Six METR staff have access comparable to Anthropic’s risk assessment teams, but customer data and model architecture remain excluded.
Anthropic said on October 15 that it had granted nonprofit evaluator METR access to internal systems comparable to that of its own risk assessment teams. The Claude developer is also negotiating with other potential evaluators, including Hugging Face.
Six staff members from METR, or Model Evaluation and Threat Research, have Anthropic hardware and workspace at its offices, the company said. They cannot access customer data, legally privileged material or details of model architecture.
Anthropic said it has no veto over METR’s conclusions about risks, incidents or company practices, though it may demand redactions within narrowly specified limits. METR may publicly challenge those restrictions without disclosing the information withheld. The nonprofit said its report would explain the boundaries of its work and access.
The arrangement falls within the highest of three access levels Anthropic is developing. At the lowest level, reviewers undertake limited investigations and require permission to publish. The middle level offers additional information and allows publication without prior approval, subject to redactions that reviewers may contest publicly. METR’s level adds employee-like access while retaining similar publication rights.
Anthropic said its other evaluation partnerships remain at an early stage, with narrower access and publishing rights. Biosecurity nonprofit SecureBio is in the lowest tier and is examining the chemical and biological weapons section of Anthropic’s August risk report. The company said it would consider broader rights after that review.
Anthropic is also working with Faculty, Accenture’s AI business, on model evaluations, adversarial testing and checks on whether models behave as intended. Experts have questioned evaluator independence because Anthropic and OpenAI select the organizations that review their systems.