Frontier AI companies have already pledged to use risk thresholds, governance and external evaluations to manage severe model risks. A new White House accord makes outside evaluation one of four named layers of controls and audits and adds an independent board committee.
Executives at Google, Anthropic, Meta, OpenAI, xAI and Nvidia signed the “White House Accord on Super Intelligence: Joint Commitment on Frontier Responsibilities.” The agreement says companies training and deploying frontier models should maintain internal controls, empower an internal team to check them, use an independent external auditor or evaluator and designate an independent board committee.
Those controls cover model capabilities and alignment during training and deployment in areas including cybersecurity, biosecurity and chemical threats. They are also intended to prevent models from hacking or accessing technical systems in unintended ways.
The outside evaluator is meant to assess whether the controls and monitoring systems work as intended, while the board committee receives reports and oversees remediation.
The accord is voluntary. It says participating companies will meet regularly to establish safety standards and best practices and leaves open the possibility that the measures could later be written into laws or regulations.
Outside audits become explicit
The six companies were already covered by the 2024 Frontier AI Safety Commitments agreed at the Seoul AI Summit. Google, Anthropic, Meta, OpenAI and xAI were original signatories, while Nvidia joined in February 2025.
The Seoul commitments also contemplated independent third-party evaluations, but “as appropriate” alongside evaluations by governments and other bodies. They went further on risk thresholds: companies agreed to define thresholds at which severe risks would be intolerable and, in the extreme, not develop or deploy a model if mitigations could not keep risk below those thresholds.
The White House accord does not repeat that stop-development language. Instead, it is more specific about who checks safeguards: an internal team, an independent external evaluator and an independent board committee. It also does not reproduce Seoul’s commitment to provide public transparency, subject to security and commercial-sensitivity exceptions.
Audit details are still open
The accord says its four layers are intended to give companies, customers and the public confidence that frontier systems are operating as intended. But it does not define who qualifies as independent, how much system access an evaluator receives, when an evaluation must happen or whether customers will see the findings.
Those details affect what an evaluation establishes. In May 2026, OpenAI said third-party evaluation reports should describe the model and safeguards tested, the tools and evaluation setup used, the resources available to the system and the checks used to validate the result. OpenAI is one of the six signatories, but those reporting requirements do not appear in the White House accord.
The companies have committed to keep meeting on standards and best practices. For now, the accord identifies who should review frontier-model controls, but not a common test those reviewers must apply.