Anthropic and Accenture pledge $2bn for embedded AI safety evaluation