Anthropic and Accenture are expanding their collaboration on artificial intelligence safety through a new embedded evaluation program for frontier AI, with each company expecting to invest at least $1 billion in the area over the next five years. The effort is aimed at a gap the industry has yet to close: making a frontier lab's safety commitments independently verifiable.

The program will be led by Faculty, Accenture's specialist AI business, which the company announced plans to acquire earlier this year as part of a broader expansion of its AI capabilities. Faculty's evaluators will work alongside Anthropic teams to evaluate and red team models, conduct alignment assessments, and test model safeguards.

Unlike conventional external evaluators, the embedded teams are expected to receive access comparable to that of Anthropic employees. This would allow them to observe models during training, examine decisions surrounding development and deployment, and assess whether the company is meeting its safety commitments. The depth of access is what distinguishes the model, bringing evaluation inside the development process rather than leaving it at the boundary of the lab.

The initiative follows a commitment made by Anthropic CEO Dario Amodei in his recent essay, "We Must Pace the Frontier," in which he called for independent evaluators to operate inside frontier AI labs as part of broader efforts to make safety commitments verifiable. The new program turns that proposal into a working arrangement, with a named partner and a defined funding commitment behind it.

Faculty has previously worked with major AI developers on safety testing, including contributing to biosecurity evaluations for Anthropic's Claude 4 models and red teaming OpenAI's o1. That track record means the team arrives with hands-on experience testing frontier models from more than one lab.

Anthropic said the framework remains experimental, as no established industry standards yet determine what embedded evaluators should be allowed to access or how their findings should be reported. In the absence of agreed norms, the program's own practices on access and disclosure will define its terms. Funding also remains unresolved. Anthropic has previously argued that independent evaluations should eventually be financed through pooled or government-backed mechanisms, but for the Accenture partnership, the company will directly fund the evaluation work, meaning the model rests for now on direct corporate funding rather than the shared infrastructure Anthropic has advocated.

The arrangement is not exclusive. Anthropic said it is also in discussions with nonprofit evaluator METR and plans to announce additional evaluation partners in the coming weeks, while Accenture will remain free to work with other AI developers. The structure points toward a roster of embedded evaluators at Anthropic rather than dependence on a single relationship.

Anthropic emphasized that embedded evaluators do not transfer responsibility for model safety to outside organizations. Instead, the program is intended to make the company's own safety commitments easier to independently verify. The coming weeks should clarify at least two things: which additional evaluators join the program, and how its experimental access and reporting terms take shape.