OpenAI will require approval for services offering direct access to its models

Services offering direct access to OpenAI models will need the company’s review and approval before providing that access under updated Usage Policies effective November 20, 2026. The requirement explicitly covers model routers and AI gateways. It puts those services under an approval requirement as OpenAI seeks to make safeguards consistent wherever customers access its models.

OpenAI announced the change on October 21, saying it had spent the past few months working with customers and partners to align safety policies. The update also aims to clarify rules for safe model development and sensitive uses of AI, alongside the new requirement for services that provide direct model access.

Safety duties follow the model across providers

The safety-policy alignment reaches beyond services subject to the new approval rule. OpenAI says its policies apply to access through OpenAI itself, partners that host its models, and reseller partners. It is working with each partner to ensure its implementation meets or exceeds OpenAI’s safety standards, while preserving customers’ choice of service.

Partners providing access to OpenAI models must carry out several linked responsibilities:

• Maintain safeguards consistent with OpenAI’s own protections.
• Investigate potential policy violations and take appropriate action when concerns are identified.
• Coordinate enforcement so restrictions remain effective across services providing access to the models.

OpenAI’s explanation distinguishes protections inside a model from those supplied by the service around it. Before deployment, the company says it evaluates capabilities and risks, trains models to follow safety requirements, and tests attempts to bypass safeguards. After deployment, it monitors for potential abuse and investigates concerns.

For cyber misuse, those layers include training models to refuse harmful requests and looking for suspicious patterns across interactions. OpenAI also points to trusted access programs, which set eligibility requirements and conditions for access to its most powerful capabilities. Its position is that both model-level protections and surrounding systems must remain effective through partners.

Some safeguards are built into the model and others depend on the systems and people around it.

OpenAI, October 21 policy announcement

Investigations without learning users’ identities

The announcement also addresses how OpenAI intends to investigate misuse without unnecessarily disrupting other users. Customers can encounter abuse through their own end users, compromised accounts, or leaked API keys, the company says. OpenAI works with customers to understand and address those problems.

OpenAI says it is working with customers to introduce SafetyIDs. Customers can provide these identifiers to help distinguish individual end users during an investigation without OpenAI learning their identities. The intended benefit is more precise action against potential misuse, with less disruption to others. The company describes this as work underway, rather than announcing general availability.

Service differences also shape how safeguards are implemented. OpenAI cites its work with customers and partners on Private Safety Processing as an example of developing effective protections while giving customers greater choice and control over sensitive data. It presents that work as part of adapting safeguards to different services and customer uses.

Restrictions can be reconsidered

OpenAI says enforcement decisions rest on documented conduct and risk, with rules applied consistently and responses proportionate to the circumstances. Restrictions are not necessarily permanent: the company says it revisits them when meaningful changes in conduct, commitments, or safeguards address the original concerns. It also commits to explaining permitted and restricted uses and how customers can resolve compliance concerns.