OpenAI publishes national security principles alongside its cyber and biosecurity partnerships

What OpenAI actually shipped on July 8

The announcement has two parts: a published set of National Security Principles, and a status report on partnerships that are already live. OpenAI frames the principles as transparency into how it approaches government partnerships and national security uses of its technology, and says they apply to current and future work — including its existing agreement with the Department of War.

What makes this more than a values statement is that OpenAI names specific programs and partners. It is not describing a future intent to work with governments; it is documenting deployments that, by its own account, happened in the past month.

The access layer: Daybreak, Trusted Access, and GPT-Rosalind

The concrete structure here is a set of gated access programs. Under the Daybreak cyber defense program, OpenAI says it has established Trusted Access for Cyber partnerships with Australia, Canada, Japan, the Republic of Korea, France, Germany, Poland, the Netherlands, and EU institutions such as ENISA, plus a separate partnership with the UK government on cyber, testing, and evaluation.

Biosecurity follows the same pattern. OpenAI points to expanded trusted access to its GPT-Rosalind model for select U.S. government and allied partners supporting public health and biodefense missions. The recurring word is 'trusted access' — a tiered entitlement, not general availability. For anyone building on top of frontier models, that is the operative detail: government use is being mediated through named, restricted access rather than an open API.

OpenAI's stated rationale is defensive advantage — the claim that in cyber defense and biological security, AI can meaningfully advantage defenders. The partner list reads as an attempt to make that specific to allied democracies rather than a generic capability.

Three restrictions stated as contract terms

OpenAI restates three limits it first articulated with its Department of War agreement: no use of its technology for mass domestic surveillance, no use to direct autonomous weapons systems, and no use for high-stakes automated decisions. It says these are consistent with the principles published today.

These are written as usage restrictions on contracts, not model-level guarantees. That distinction matters in practice — the enforcement mechanism is the agreement and OpenAI's willingness to hold to it, not a technical control a customer can inspect. The company is asking to be judged on transparency about what its technology can and cannot be used for.

Where OpenAI draws the line on its own authority

The post is explicit that OpenAI does not want to be the final decision-maker. It describes engaging national security expert David Kris to facilitate the process and provide independent judgment, holding internal listening sessions, and drawing on teams from research and safety to policy and partnerships.

Many of the most consequential questions about AI in government, especially in national security, should be answered through the democratic process. The role of companies like ours is to help inform those decisions, not make them alone.Montana Labs

On that basis OpenAI says it supports legislation to establish safeguards around the highest-risk military uses — the same three areas its contracts already restrict: domestic surveillance, autonomous weapons, and other high-risk uses.

The implication: a company's usage terms are standing in for policy that doesn't exist yet

The through-line of this announcement is that OpenAI is deploying frontier models into cyber defense and biodefense missions across roughly a dozen governments now, while the legislative safeguards it says it wants are still hypothetical. In the gap, the operative rules are OpenAI's own principles and its three contractual red lines.

For applied teams, the practical takeaway is that national security access to these models is being governed at the vendor layer — through trusted-access tiers and contract restrictions — rather than through law or a technical enforcement surface. OpenAI is candid that it would rather the democratic process set those limits; until it does, the company's published principles are the binding document, and they are only as strong as the disclosure and enforcement behind them.