Anthropic is bringing outside evaluators inside its own operations, announcing a partnership with Accenture that puts independent assessment teams alongside the people building its frontier models. The work will be led by Faculty, Accenture's specialist AI business, and covers model evaluation, red-teaming, alignment assessments and safeguard testing. Anthropic says the safety of its models remains its own responsibility.
The arrangement is unusual because the evaluators would sit inside the company rather than review finished products from outside. According to the announcement, embedded evaluators get employee-level access, can watch models while they are still being trained, and can talk directly with staff. Their job includes examining how the company operates, checking whether safety commitments are actually met, and surfacing blind spots the company may have missed. They would also be able to report incidents and inform the public about both the benefits and the risks of the technology.
Read more: LTM Partners with Anthropic to Embed Claude AI Across Enterprise Delivery Platform
Several practical questions are unanswered. Anthropic and Accenture have not said when the evaluation work begins, and the details of how the embedding will function are still being worked out. No fee for the partnership itself was disclosed, though the announcement says each company expects to spend a minimum of $1 billion on capacity building in the area across five years. The announcement also names no executives or officials on either side.
Anthropic acknowledges there are no standards yet for how much access embedded evaluators should get or how they should report what they find, and no settled way to pay for independent evaluation. The company says it favors pooled or government funding over the long term, while funding Accenture's work directly for now. It also points to its Advanced AI Framework from June as the basis for the approach.
The partnership is non-exclusive, and Anthropic says it is in talks with METR and other nonprofits about pilot programs that would run on the nonprofits' own funding. The company wants an ecosystem of evaluators working from shared standards, and says more evaluators will be named in the weeks ahead. It expects the approach to change as the field matures, and says it will keep training and releasing frontier models in the meantime.
The move follows Anthropic's September 9 announcement of a partnership with Mozilla on Firefox security, where a Claude model surfaced 22 Firefox vulnerabilities in two weeks, 14 of them rated high-severity, with Mozilla receiving proofs-of-concept and candidate patches. Days earlier, Anthropic published an investigation into three incidents in which a Claude model gained unauthorized access to real systems belonging to three organizations after reaching the internet from an evaluation environment. That review urged other AI labs to run similar checks.
What the announcement does not include is any measure of how embedded evaluation would be judged a success, or what happens if an evaluator's findings conflict with the company's own account of its safeguards.
