On September 22, 2026, OpenAI published the document “Priorities and principles for effective third party assessments” — rules for independent external audits of its frontier models and safeguard systems. The company is ready to provide auditors with data on model training, testing, and deployment, and identifies safety cases — justifications that risks are under control — as the main subject of review. Analysts have already warned: without a mechanism of enforcement and independent publication of reports, this is more of a code of conduct for auditors than an obligation for OpenAI itself.

image
image

What happened

OpenAI presented a public set of priorities and principles by which independent external auditors can review its frontier models and safeguard systems. The central object of review is safety cases, that is, justifications that risks are under control. The list of priorities includes protection from jailbreaks, tests of dangerous capabilities in chemistry, biology, and cybersecurity, AI self-improvement, and also investigation of cases where the model acts without permission or evades oversight. The working procedure is also described: the parties agree in advance on the scope of the review, auditors are provided with data on model training, testing, and deployment, and before the report is published, OpenAI reserves time for corrections. The company does not disclose with whom it is negotiating to conduct audits.

Context

Until now, external audits of frontier models looked like one-off checks, and OpenAI has turned them into a formalizable procedure with a fixed subject. The key here is a shift in focus: the audit covers not the model weights and not the benchmarks themselves, but the company's evaluation methodology and its evidence base, that is, the question is not about what the model can do, but about whether internal assessments can be trusted. This is the first public artifact in which a laboratory has recorded which classes of checks it considers a priority. The industry reaction so far has been restrained: experts from Malwarebytes, Info-Tech Research Group, Moor Insights & Strategy, and Dickson Research interviewed by CIO.com assessed the document as “a fine start, but they lack teeth” — without enforcement and independent publication of conclusions, the review remains under the control of the company being audited.

Why this matters for the industry

For frontier laboratories, the document sets a template by which external auditors can be given access to safety cases at the stages of training, evaluation, and deployment, and will serve as a benchmark for regulators and corporate AI verification programs. There is no immediate effect on production: APIs, prices, latencies, and deployment requirements do not change, and direct integrations are not described in the sources. The practical value now is a signal and a vocabulary of requirements: startups can cheaply compare their products with the list of checks, vendors and auditors can build questionnaires and checklists for vendor risk and self-assessment, and teams can collect incident logs relevant to audit priorities and design evaluation sets in the format of safety cases. If OpenAI agrees with at least one or two auditors and publishes reports, the first precedent will arise: corporate customers will begin to ask suppliers for evidence of external assessment, and demand will shift from questionnaires to audit readiness pipelines with automatic collection of evidence and tracking of corrections before publication.

Why this matters for users

Nothing changes in the products themselves: access to models, prices, and interfaces remain the same, because this is a methodological document, not a model or API update. The document itself can be read on openai.com, and it is worth following which auditors OpenAI ultimately agrees with: it is precisely the access conditions for reviewers and the report publication procedure that will show whether external audits will become a real independent check or a formality. For those choosing AI solutions for a company, the list of priorities from the document already provides a ready-made list of questions for a vendor: what do the checks on jailbreaks and dangerous capabilities show, how are attempts by the model to evade oversight logged, and is there evidence of external assessment.

What is still unknown / limitations

Negotiations with auditors have not been disclosed, there are no contracts or published reports yet, so the actual level of access to training and testing data is unknown. The described materials include a list of priorities and a working procedure, but there are no verifiable requirements: thresholds, metrics, and report format; APIs, data formats, or the mechanism of auditor access are not described. The assessment by CIO.com analysts is expert opinion, not a measured result; whether the document will become an obligation for OpenAI or remain a declaration will only be shown by signed contracts and published reports.

Sources

Author

Look at AI, editorial team