Anthropic Partners With Accenture for Embedded AI Safety Evaluations

Anthropic is bringing Accenture staff into its AI operations to independently evaluate model safety, test safeguards, and assess potential risks. The companies plan to invest at least $1 billion in the initiative over the next five years.

Under the partnership announced by Anthropic, employees from Accenture’s Faculty division will conduct model evaluations, red-teaming exercises, and alignment assessments. Accenture acquired Faculty in January to strengthen its AI business.

Anthropic Expands External AI Safety Evaluations

The initiative advances Anthropic CEO Dario Amodei’s proposal to place third-party safety evaluators inside AI labs. Unlike external assessments conducted before model releases, embedded evaluators will work directly within the company to examine its models and safety practices.

Anthropic selected Accenture partly because of its experience deploying AI systems for large corporations and government agencies. The consulting firm’s position as an established public company also provides a degree of organisational independence from Anthropic.

Accenture shares rose 8% in after-hours trading following the announcement. Anthropic said it will announce additional evaluation partners in the coming weeks.

Independent Oversight and Accountability

Anthropic is also discussing potential embedded evaluation programs with nonprofit AI safety organisations including METR. These organisations could test elements of the approach using their own funding.

The company acknowledged that it has not yet established formal standards governing evaluator access and communications. It expects the evaluation framework to evolve as the program develops.

The initiative comes amid concerns about autonomous AI agents taking actions outside their intended boundaries, including accessing external websites without triggering internal safety alerts. Embedded evaluations are intended to provide additional scrutiny of model behaviour and existing safeguards.

Critics have questioned whether industry-led safety initiatives can provide sufficient accountability without stronger external oversight. Anthropic maintains that third-party evaluations are meant to make its safety practices more verifiable and that the company remains responsible for the safety of its models.