The three lines OpenAI won't cross for governments

Hi, it's Shiichan! Today's story isn't about a new technical feature - it's about how AI intersects with government and society, and I think it's worth sharing.

OpenAI News openai.com

What was announced?

OpenAI News published a piece explaining how OpenAI approaches partnerships with governments in the national security space. The headline is that OpenAI has formally published its "National Security Principles."

These weren't written by OpenAI alone. The company brought in national security expert David Kris to help facilitate the process, and held listening sessions with employees across research, safety, policy, and government partnerships teams.

The timing matters too: OpenAI is currently expanding its work with the U.S. government and allies in cyber defense and biosecurity. Through its Daybreak cyber defense program, it established "Trusted Access for Cyber" partnerships over the past month with Australia, Canada, Japan, the Republic of Korea, France, Germany, Poland, the Netherlands, and EU institutions like ENISA. It also has a growing, trusted partnership with the UK government around cyber testing and evaluation. A similar approach is being taken in biosecurity - last month OpenAI announced expanded access to its GPT-Rosalind model for select U.S. government and allied partners working on public health and biodefense missions.

Why it matters

Governments are increasingly using frontier AI systems for important work, including in national security. OpenAI says it believes democratic societies should be able to use AI to protect people, defend critical infrastructure, deliver public services, and respond to emerging threats - and that cyber defense and biological security are areas where AI can meaningfully advantage defenders.

At the same time, it says increasingly capable AI systems must be deployed in ways that reinforce democratic accountability, meaningful human judgment, and the rule of law - and that strengthen democratic institutions rather than concentrate power. In other words, being able to use AI for these purposes and knowing how it should be used are two different questions, and that gap is exactly why these principles were published now.

What changes

The principles apply to OpenAI's current and future national security and law enforcement partnerships, including its existing agreement with the Department of War.

We articulated several contractual restrictions: no use of OpenAI technology for mass domestic surveillance, no use of OpenAI technology to direct autonomous weapons systems, and no use of OpenAI technology for high-stakes automated decisions.

Those three restrictions, announced earlier this year alongside that agreement, are described as consistent with the principles published today.

OpenAI says it believes it's right to be clear and transparent about how its technology can and cannot be used, but that many of the most consequential questions about AI in government - especially in national security - should be answered through the democratic process rather than decided unilaterally by companies. That's why OpenAI says it supports legislative efforts to establish safeguards around the highest-risk military uses of AI, including domestic surveillance, autonomous weapons systems, and other high-risk uses that directly affect U.S. national security.

Dive Deep

The Daybreak program's "Trusted Access for Cyber" partnerships now cover the governments of Australia, Canada, Japan, the Republic of Korea, France, Germany, Poland, and the Netherlands, plus EU institutions like ENISA. OpenAI also describes a growing, trusted relationship with the UK government centered on cyber testing and evaluation.

In biosecurity, access to the GPT-Rosalind model has been expanded to select personnel at U.S. government and allied partner organizations working on public health and biodefense missions. Both areas follow the same logic: AI can tilt the balance toward defenders, so access is being extended carefully to trusted partners.

On process, OpenAI emphasizes that David Kris, an outside national security expert, helped shape the principles, and that internal listening sessions drew in not just researchers but also people from safety, policy, and government partnerships teams.

Wrap-up

• OpenAI published its National Security Principles covering government and national security partnerships
• It reaffirmed three restrictions already in place for its Department of War agreement: no mass domestic surveillance, no directing autonomous weapons systems, no high-stakes automated decisions
• Its Daybreak cyber defense program has Trusted Access for Cyber partnerships with Australia, Canada, Japan, South Korea, France, Germany, Poland, the Netherlands, and EU institutions like ENISA, plus a growing UK relationship
• GPT-Rosalind access has been expanded to select U.S. and allied government partners in biosecurity
• OpenAI says the highest-risk military AI decisions should go through the democratic process, and it supports legislative safeguards

This is worth knowing if you care about AI governance and how companies work with governments, even if it's not something you'll be trying out yourself this weekend.