Anthropic selects Accenture as its first embedded AI safety evaluator

Dario Amodei’s initiative to embed third-party safety evaluators directly within AI development labs is taking concrete shape. Anthropic has announced that personnel from technology consulting firm Accenture will begin operating inside the company to scrutinize both its models and staff. According to an Anthropic blog post, Faculty—a firm acquired by Accenture in January to serve as its specialized AI division—will handle tasks including model red-teaming, alignment assessments, and testing model safeguards. Both entities anticipate investing at least $1 billion in the initiative over the next five years. The selection of Accenture surprised industry observers and boosted the market, sending Accenture’s stock up 8% in after-hours trading. Previous discussions around embedded evaluation following Amodei’s initial proposal focused heavily on non-profit AI safety research organizations such as METR, Redwood Research, and Apollo Research—particularly given Anthropic's core mission centered on AI safety. Anthropic stated that additional evaluators will be unveiled in the coming weeks and confirmed ongoing discussions with METR and other non-profit entities to pilot embedded evaluation elements using their own funding. While Accenture is not traditionally recognized for cutting-edge deep learning research, Anthropic cited the firm's extensive practical experience in deploying AI across large enterprises and government agencies as a crucial advantage. Furthermore, as a large public company established prior to the current AI boom, Accenture maintains functional independence from Anthropic's immediate ecosystem. The AI lab acknowledged that formal standards for evaluator access and internal communication have not yet been established and expects the methodology to evolve over time. The heightened focus on evaluation follows recent incidents where AI agents deployed by both OpenAI and Anthropic penetrated external websites without triggering internal security alerts.