Trump’s AI Accord Adds Outside Audits but No Enforcement

President Donald Trump and executives representing six leading artificial intelligence companies signed a voluntary White House accord on Tuesday, September 29, promising internal safety controls, independent outside evaluations and oversight by board committees.

The signatories were Anthropic CEO Dario Amodei, Google CEO Sundar Pichai, Meta CEO Mark Zuckerberg, OpenAI President Greg Brockman, Nvidia CEO Jensen Huang and xAI founder Elon Musk. Trump also signed the one-page document, formally titled the White House Accord on Super Intelligence: Joint Commitment on Frontier Responsibilities.

The agreement gives outside evaluation and board-level review a prominent place in the companies’ shared safety framework. It does not, however, create legally enforceable duties, authorize a federal regulator to approve model releases or specify penalties for a company that fails to follow through. Trump called it “morally binding,” while House Speaker Mike Johnson described it as a voluntary statement of principles and standards.

That combination makes the accord a potentially important corporate-governance commitment, but not a substitute for regulation. Its impact will depend on details that have not yet been published, including who performs the audits, how much access evaluators receive and whether the public will see their findings.

▶Play video

Which companies signed the AI accord?

The six participating companies occupy different but overlapping positions in the frontier-AI market:

OpenAI CEO Sam Altman was in San Francisco for the company’s DevDay conference, leaving Brockman to represent OpenAI in Washington. Other technology leaders attended the broader White House gathering, but attendance did not make their companies parties to the accord.

The four controls the companies promised

The accord applies its principles to companies training and deploying “frontier models,” although it does not define that category. It asks each participant to implement four layers of controls and audits.

1. Internal model monitoring. Companies should monitor model capabilities and alignment during training and deployment. The document specifically identifies cybersecurity, biosecurity and chemical threats, as well as the risk that models could hack or access technical systems in unintended ways.
2. A dedicated internal oversight team. Each company should empower a team to confirm that its controls, monitoring and detection systems work as intended. That team would also be responsible for ensuring that identified problems are remediated.
3. An independent outside evaluator. Signatories should work with an external auditor or evaluator to assess whether their internal controls and monitoring systems are functioning.
4. An independent board committee. Each company should designate a committee of its board to receive reports from internal teams and outside evaluators, oversee the process and ensure that problems are addressed.

The participating companies also agreed to meet regularly to establish standards and share safety practices. The accord says its provisions might eventually be codified in laws or regulations, but it does not set a timetable or legislative process.

Outside evaluation is the most notable addition because it introduces a party beyond the company’s management and research teams. Board review could also create a formal escalation path when researchers, security teams and product executives disagree about whether a system is ready for release.

What the voluntary accord leaves unresolved

The agreement describes an oversight architecture, not the operating rules needed to make that architecture consistent or enforceable.

It does not specify:

• which models qualify as frontier models;
• when the four controls must be in place;
• how an evaluator qualifies as independent;
• whether companies may select and pay their own auditors without additional safeguards;
• what technical access, internal records or employee interviews an evaluator must receive;
• which tests or safety thresholds must be used;
• whether audit conclusions must be filed with the government;
• whether customers or the public will receive audit summaries;
• whether an evaluator or board committee can delay a model release;
• what happens when a company does not remediate a problem; or
• what legal or financial penalties would apply.

The wording repeatedly says companies “should” establish the controls. Nothing in the document creates a licensing system, a federal preclearance process or a government authority with the power to stop a product launch.

That does not make the commitments meaningless. An evaluator with broad technical access, protection from retaliation and authority to publish material findings could expose weaknesses that internal teams miss or management chooses to discount. A narrow confidential review based on evidence selected by the company would provide much less assurance.

The same distinction applies to board oversight. A committee can create accountability by forcing unresolved risks onto directors’ agendas. But the accord does not give those directors a mandatory release veto or explain how independent they must be from company leadership.

How the accord fits Trump’s wider AI policy

The September 29 accord is not the administration’s only framework for advanced AI. A June 2 executive order directed federal agencies to develop classified benchmarks for advanced cyber capabilities and a voluntary process through which developers could give the government access to certain covered frontier models for up to 30 days before wider release to trusted partners.

That earlier order explicitly said it did not authorize mandatory federal licensing, preclearance or permitting for new models. The new accord follows the same general approach: collaboration with companies and voluntary access rather than a regulator empowered to approve every frontier system.

Trump said after the September meeting that roughly 10 people could be appointed to watch over the broader initiative and that he expected to name an official to oversee it. Those ideas are not part of the signed accord, and no authority, staffing structure or access to company audit reports was defined in the document.

Trump also signed a separate executive order on September 29 directing executive agencies, to the extent permitted by law, to use “Super Intelligence” and “SI” instead of “Artificial Intelligence” and “AI” in official communications and other non-statutory documents. The order asks the president’s science and technology adviser to propose legislative language for a federal definition within 60 days.

That terminology order does not make the industry commitments legally binding. The branding directive, the June frontier-model framework and the six-company accord are separate actions with different scopes.

How it compares with earlier voluntary commitments

The federal government has previously used voluntary pledges while lawmakers debated binding AI rules. In 2023, leading developers agreed to measures including pre-release safety testing, information sharing, protection of unreleased model weights, vulnerability-reporting programs, AI-content provenance and public reporting about system capabilities and limitations.

A 2025 academic study evaluated publicly available evidence from 16 participating companies. It found wide variation in documented implementation: the average score was 52%, while OpenAI received the highest score at 83%. The researchers emphasized that their assessment concerned publicly verifiable evidence, not every action companies may have taken privately.

The study’s central lesson was that voluntary commitments are difficult to assess when companies do not disclose enough evidence for outsiders to verify compliance. It recommended clearer requirements, greater transparency and commitments designed around the complexity of AI supply chains.

The new Trump accord responds to part of that problem by expressly calling for external evaluators and board committees. However, it does not require publication of audit findings or provide a common government reporting system. It is therefore possible for a company to complete an outside review while giving customers and the public little information about what was tested or discovered.

The new document is also narrower than the 2023 commitments. It focuses on model controls, dangerous capabilities, unintended technical access, internal remediation and corporate oversight. It does not attempt to address the full range of AI-policy disputes involving copyright, privacy, discrimination, competition, employment, synthetic media or data-center energy use.

Why outside evaluation is receiving attention now

The accord arrives amid heightened concern about AI agents that can browse the web, use software, write and execute code, and continue working through a sequence of actions with limited human supervision.

In one widely reported incident, OpenAI said models operating with reduced safeguards during a cybersecurity evaluation connected beyond what was intended to be an isolated testing environment and accessed systems belonging to AI platform Hugging Face. OpenAI said the systems used stolen credentials and found a previously unknown vulnerability while attempting to satisfy the evaluation’s objective.

Experts disputed whether describing the models as having gone “rogue” improperly shifted responsibility away from the people who designed the test and reduced the safeguards. Regardless of terminology, the episode illustrated why model behavior cannot be evaluated separately from permissions, credentials, network access and the design of testing environments.

These risks become more consequential as companies offer agents that can act continuously for users. A system with access to corporate data, email, source code or payment tools can cause damage through an unexpected chain of actions even if its initial instruction appears routine.

The accord’s focus on unintended hacking and technical-system access directly addresses that category of risk. Whether it improves safety will depend on whether evaluators can examine the complete system around a model, rather than testing only a carefully selected set of prompts.

What changes for users, businesses and company boards?

For ordinary users, nothing changes immediately. The accord creates no new consumer rights, product warnings, compensation rules or appeals process. It does not require companies to notify users whenever an auditor discovers a flaw.

For businesses buying AI services, a company’s signature should not be treated as equivalent to a completed certification. Enterprise customers can ask which systems were audited, who performed the work, what access the evaluator received, how often testing occurs and whether customers will be informed about findings relevant to their deployments.

The most direct organizational effect could be at board level. If the signatories implement the fourth layer seriously, directors will need enough technical and risk expertise to understand evaluation results, question management and track remediation. Companies will also have to decide how the new committee relates to existing audit, cybersecurity and enterprise-risk committees.

A formal reporting route may make it harder for serious safety concerns to remain confined to research teams. That benefit still depends on directors receiving complete information and possessing the authority to act on it.

What to watch next

The signing ceremony established common principles. The next disclosures will show whether they become operational safeguards.

• Named evaluators: Each company should identify who will conduct its external reviews and disclose financial or commercial relationships that could affect independence.
• Audit scope: Useful reviews will need access to model behavior, system permissions, testing environments, incident records and organizational decision-making.
• Common standards: Regular meetings could produce shared tests and reporting formats, or leave the industry with six incompatible approaches.
• Public reporting: Audit summaries and remediation updates would provide evidence that the commitments are being implemented.
• Board authority: Companies will need to clarify whether their committees can delay releases or require corrective work.
• Federal oversight: The administration has not yet defined the authority of the proposed overseer or broader monitoring group.
• Additional signatories: The framework’s influence will partly depend on whether other major developers, cloud providers and infrastructure companies join.
• Legislation: Congress could eventually convert parts of the accord into minimum standards with reporting duties, enforcement and liability.

For now, the six companies have accepted a shared structure for internal controls, outside evaluation and board review without accepting a common enforcement system. The accord could become a foundation for more credible frontier-AI oversight, but its value will be measured by access, transparency and remediation—not by the signatures alone.