Hi
Welcome (back) to The Prompt. We’re recapping recent updates, sharing what’s top of mind for us, and calling out favorite ways in which everyday people are using our tools.
If this was forwarded to you and you find it helpful, make sure you’re signed up for the next issue.
[Policy] Making AI audits and assessments work
We’re outlining the principles we believe should guide a strong system of independent AI safety and security audits and independent third party assessments. Safety and security should be central to how AI is built and governed. Independent review can help confirm that companies are following the law, meeting their commitments, and addressing meaningful risks. Within this system, audits and independent third party assessments play distinct but complementary roles: audits examine whether an organization follows established requirements and processes, while third party assessments test whether evidence supports specific safety or security claims.
OpenAI has consistently supported independent third-party assessments tied to clear safety goals. At the state level, we have supported independent audits of youth safety protections through the Parents & Kids Safe AI Act and audits of frontier safety frameworks in Illinois’s SB 315. At the federal level, our Frontier Safety Blueprint calls for the federal government to require annual independent audits of frontier developers’ safety frameworks and independent third party assessments of the most advanced models, including to monitor progress towards recursive self-improvement.
Our commitment to both audits and independent third-party assessments extends beyond formal legal requirements. OpenAI has long collaborated with organizations including METR, SecureBio, Apollo Research, and Irregular to evaluate frontier capabilities and risks such as long-horizon autonomy, biological threats, cybersecurity, scheming, deception, and oversight subversion. These assessments supplement our internal testing, help identify blind spots, and provide additional evidence to inform deployment decisions.
Requiring companies to undergo review is not enough. A defined control framework is a prerequisite for effective audits. The enclosed six principles describe how we can make AI auditing and independent third party assessments work by building an ecosystem that is independent, qualified, technically informed, capable of identifying material findings, and accountable for the quality of its own work.
1. Define the audit’s scope and match requirements to risk and responsibility.
An audit or assessment should begin with a clearly defined scope: the systems, products, environments, activities, and controls under review. Within that scope, auditors should identify the relevant safety or security objectives, legal requirements, and assessment criteria. For children, that might mean assessing whether age-appropriate protections are in place and effective. For frontier AI, it might mean assessing catastrophic-risk controls, internal governance, or the security measures protecting models and sensitive infrastructure. Different objectives require different evidence, expertise, and assessment methods.
Requirements should also be tailored to each entity’s capabilities, responsibilities, and control. Obligations developed for one risk domain should not automatically extend to differently situated actors, such as smaller developers, researchers, startups, downstream users, or participants in the open-model ecosystem.
2. Use the right kind of review
Audits and independent third party assessments serve different but complementary purposes. Audits examine whether an organization follows applicable requirements, governance processes, and safety or security commitments. Independent third party assessments examine whether evidence supports a specific safety or security claim, such as whether an AI system, safeguard, or security control is fit for its intended purpose.
For example, an audit might examine whether an advanced AI model developer followed its safety framework and documented required evaluations. A validation or verification assessment might examine whether the underlying tests, evidence, and safeguards support claims about model capabilities or safeguard sufficiency. Both require competent reviewers, but advanced-model validation requires specialized technical expertise, secure testing environments, and, for frontier AI, national security awareness.
3. Follow common standards
Requiring companies to undergo audits or independent third-party assessments is not enough; reviewers also need a clear framework for what they are assessing and how. Common, interoperable standards should define the risks and controls in scope, the evidence and criteria reviewers should use, the level of assurance expected, and how assessments should be conducted and reported. Without that shared foundation, reviewers may apply inconsistent questions and thresholds, making findings difficult to compare or act on.
Common national or international standards should establish a consistent safety and security baseline for audits and assessments. States may adopt different protections, but comparable risks should be evaluated using rigorous, interoperable criteria and methods so findings are credible, consistent, and capable of driving system-wide safety improvements.For frontier AI, federal coordination is especially important because catastrophic risks transcend state borders and credible evaluations may require specialized expertise, secure infrastructure, and national security capabilities. For emerging risk areas that may not yet have established standards, frameworks should evolve quickly and responsibly with input from auditors, assessors, and relevant domain experts.
4. Ensure credibility and independence
Audit and independent third party assessment processes should be designed to resist both regulatory capture and politicization. Auditors and assessors should be accredited and selected through documented, consistently applied criteria, such as national and internationally recognized qualification standards, that match their competence to the scope of the review. Such criteria could include relevant training, assessment experience, knowledge of applicable requirements, and where needed, technical and sector-specific expertise to evaluate the evidence.
To protect their independence, reviewers should disclose and appropriately manage material conflicts, avoid assessing their own work, and remain free from relationships or outcome-contingent compensation that could compromise their judgment. Companies should be able to choose among qualified assessors, but the system should guard against “assessor shopping” designed to avoid or soften adverse findings.
Findings should be based on objective, verifiable evidence.. Neither the companies being reviewed nor political actors that helped shape the rules should be able to steer individual findings or outcomes.
5. Enable secure oversight and responsible transparency
Auditors, independent assessors, and authorized regulators should receive secure access to the necessary evidence they need, while user data, model weights, confidential intellectual property (IP), and security information remain protected from unnecessary disclosure, public release, or misuse. That access should not require wholesale production or public disclosure of all system logs, private chain-of-thought, model weights, personal information, trade secrets, or other security-sensitive information. Access should be proportionate to the audit’s scope and subject to appropriate confidentiality, privacy, security, data-minimization, retention, and use restrictions.
Public reporting should provide meaningful accountability without creating new risks. Plain-language summaries should identify the audit’s scope, applicable standards, methodology, assurance level, material findings and limitations, required corrective actions, and remediation status. They should also disclose the auditor’s identity, relevant qualifications, and material conflicts of interest. Regulators should receive complete reports and supporting evidence through protected channels, while public disclosures may use targeted redactions necessary to protect privacy, cybersecurity, IP, public safety, and national security.
6. Make audits and assessments lead to safer, more secure systems
Audits and independent third party assessments should lead to measurable improvements in safety and security. Material findings should result in time-bound remediation and appropriate follow-up, including verification that significant deficiencies have been corrected. Oversight bodies should use audit and assessment findings to identify systemic gaps, assess evaluator performance, and strengthen standards as technology and risks evolve. Given the evolving nature of both the capabilities and the methods to safeguard against harms, this will inherently require iterative approaches to understanding appropriate remediation steps congruent with updated standards. Where applicable, regulators should take appropriate enforcement action in response to repeated or egregious failures, material misrepresentations, or obstruction of legitimate oversight.
Audits and assessments should complement ongoing testing, incident reporting, and legal enforcement – not replace it.
[Perspective] The next threshold
The frontier of AI is extending, and the world needs to prepare for what that means.
For years, progress in AI was often measured by whether models could answer more questions, score higher on benchmarks, or perform familiar tasks more accurately. Increasingly, that is not the whole story. The frontier is extending into kinds of work that, until recently, were largely the domain of highly specialized experts: sustained reasoning, complex problem-solving, tool use, and longer sequences of actions directed toward an objective.
We recently saw a striking example in mathematics and theoretical computer science. An internal version of our next major model generated mathematical arguments on 10 long-standing problems, including results that resolved or made substantial progress on questions researchers had worked on for years. That was an important step toward giving scientists and mathematicians tools that can accelerate discovery.
The broader point is that AI is not simply getting better at the tasks it already performs. It is becoming capable of doing qualitatively different kinds of work – deeper, longer, and more sophisticated work that can increasingly resemble the efforts of highly trained experts.
That creates extraordinary possibilities. It also means that assumptions built around the limits of earlier systems won’t hold forever.
The same advances in reasoning that can make progress on difficult mathematical problems can help scientists pursue new discoveries. Systems that can perform increasingly complex work can help researchers test ideas faster, help businesses solve harder problems, and give individuals and small organizations capabilities that were once available only to large institutions. As the cost and effort required to perform sophisticated work fall, many more people and organizations will be able to attempt things that once demanded scarce expertise, large teams, or significant resources.
That is why greater capability also means greater responsibility, and why preparation matters now.
As models become more capable, the answer can’t be to just pretend the risks don’t exist, or to assume powerful capabilities can just be kept scarce. We need to measure them, publish what we learn, strengthen safeguards, improve alignment, and help institutions adapt before these capabilities become more widespread.
That is one reason why we continue to believe in iterative deployment: bringing capabilities into the world incrementally so that society can learn and adapt alongside the technology. Real-world use helps us understand how people actually use increasingly capable systems, identify problems that controlled evaluations cannot always anticipate, and improve safeguards based on evidence. Transparency matters for the same reason. Governments, researchers, businesses, and the public need a clear picture of what these systems can and cannot do so they can prepare.
There is another reason to keep extending access to these capabilities: resilience.
Powerful technologies don’t remain scarce forever. AI capabilities will continue to diffuse across countries, companies, and open models. In that world, safety and security cannot depend solely on trying to prevent access. They must also depend on making sure people and institutions have powerful tools of their own, can adapt quickly, and can strengthen the systems around them as capabilities spread.
That means putting advanced AI in the hands of scientists, businesses, governments, educators, and everyday people who can use it to solve problems, find weaknesses, accelerate discovery, and improve the resilience of the institutions and systems we rely on. With ChatGPT for Academic Researchers, we’re putting our frontier models and tools in the hands of 100,000 scientists, mathematicians, and engineers, at no cost.
The frontier is extending. The important question is not whether these capabilities will spread, but whether people and institutions will be ready when they do. There will be risks, and there will be mistakes. The answer is not to stop moving forward – it’s to move forward with our eyes open: measuring what these systems can do, being transparent about what we find, strengthening safeguards as capabilities grow, and giving people the tools they need to adapt and build a more resilient world.
That is realistic optimism: taking the risks seriously enough to prepare for them, while moving quickly enough to capture what these capabilities can make possible.
[Weekly wrap] This week at OpenAI
Use case we love: Tami and Ted Taylor are using ChatGPT to help run three New Mexico restaurants, from keeping up with health requirements and responding to customers to analyzing prices and drafting employee policies. AI is giving them access to capabilities that once would have required outside specialists.
Building healthy AI habits for young people: OpenAI and the American Psychological Association are partnering to develop practical guidance for families, clinicians, and educators on how young people use AI, with a focus on healthy use, age-appropriate experiences, and keeping human relationships at the center.
Better AI, available to more people: We updated GPT-5.6 Sol in ChatGPT to make answers more focused and reliable, while expanding GPT-5.6 Luna to Free users with unlimited text chats and a new Think option for harder questions. The goal: make our newest capabilities useful to more people.
AI meets people where healthcare begins: Inc examines how people are increasingly turning to ChatGPT for health and wellness questions, particularly outside normal clinic hours, and how AI can help patients navigate an overburdened healthcare system.
[Disclosure]
Graphics created by Base Three using ChatGPT.
Yawn I could make it more exciting if you want to? with 6 words not paragraphs. Always down for a Debate on AI ethics♾️💪🦾⏳️🕳🥘🎨🚢🕳🍿🍿🍿🍿🍿🍿 🕳🪨🪨🪨🪨🪨🪨
Audits and assessments only mean something if the entity being measured doesn’t also control the measurement’s inputs. Otherwise it’s an issuer-paid rating dressed up as independent review.