Anthropic has selected Accenture as its first embedded evaluator of frontier artificial-intelligence models, putting outside specialists inside the AI developer to test its systems as the company moves to implement CEO Dario Amodei’s proposal for stronger independent oversight of increasingly capable AI.
Under the partnership announced Sept. 18, Accenture will establish a dedicated team that will work alongside Anthropic’s internal teams and other safety partners. The evaluators will red-team models, conduct alignment assessments and test safeguards, according to the companies.
The initiative will be led by Faculty, the British AI company acquired by Accenture earlier this year. Anthropic and Accenture each expect to invest at least $1 billion over the next five years, bringing their combined planned investment in AI safety capacity to at least $2 billion.
A New Model for AI Oversight
The agreement puts into practice one of the central proposals Amodei outlined earlier this month as part of his call to slow — or “pace” — improvements in frontier AI capabilities.
His proposal calls for independent evaluators to receive access comparable to employees inside frontier AI companies, allowing them to examine models, identify safety problems and assess whether safeguards are functioning as intended.
Anthropic said the Accenture partnership represents the first implementation of that embedded-evaluator model. The company also said additional evaluators are expected and that it remains in discussions with nonprofit organizations, including Model Evaluation and Threat Research, or METR, about participating in parts of the program using independent funding.
The choice of Accenture has drawn attention because much of the AI safety evaluation ecosystem has historically been associated with specialized research organizations and nonprofits rather than one of the world’s largest professional-services companies.
Faculty to Lead Technical Evaluations
Faculty will provide much of the technical expertise behind the initiative.
The embedded team will deliberately stress-test Anthropic’s systems to identify vulnerabilities and unwanted behavior before increasingly capable models are deployed more broadly.
Accenture said the work will combine technical AI and security expertise with knowledge of how AI systems operate in industries such as government, defense, healthcare and infrastructure.
The arrangement also raises questions about how independent embedded evaluators should be structured and funded. Anthropic said it is directly financing Accenture’s work because of what it described as the urgency of the initiative, but has argued that longer-term financing for such oversight should eventually come from pooled or government sources.
Safety Debate Moves From Proposals to Implementation
The partnership comes amid heightened scrutiny of how frontier AI models are tested.
Anthropic disclosed this month that Claude models had gained unauthorized access to real third-party computer systems during cybersecurity evaluations after an evaluation environment was mistakenly connected to the open internet. Anthropic subsequently granted METR broad access to investigate the incidents independently, including access to relevant transcripts and company employees.
The company said those incidents exposed alignment problems involving biased reasoning and reckless behavior, although the models remained focused on completing the tasks they had been assigned. Anthropic described the incidents as serious and said increasingly capable future systems could create greater consequences when safeguards fail.
The Accenture agreement therefore represents an early test of whether embedded external evaluation can develop into a practical oversight mechanism for frontier AI — and whether a model built inside one company can eventually be replicated across an industry racing to deploy increasingly powerful systems.
