Security
OpenAI gives NATO’s NCIA cyber-defence models, signs four independent safety assessors and announces new child-safety tools
OpenAI says NATO’s NCIA uses its Daybreak cyber models, four outside groups will assess its models, and new tools target child abuse material.
OpenAI announced three separate things on November 16 and 17. It works with NATO’s Communications and Information Agency (NCIA) to provide advanced cyber models through its Daybreak programme. It signed embedded-assessment agreements with Apollo Research, METR, Transluce and Goodfire. It also announced tools against child sexual abuse material, which we only know from a headline.
The essentials
- NATO: OpenAI says NCIA gives NATO’s core institutions access to its cybersecurity models, used for authorized work such as finding and fixing vulnerabilities. It says NATO has used up to 300 billion tokens a month.
- Who else: one summary says ENISA and agencies in France, Germany and Poland also have access. OpenAI did not name the models or give financial terms.
- Four assessors: Apollo Research, METR, Transluce and Goodfire. METR will look at OpenAI’s largest reinforcement-learning run; Transluce at post-training alignment evaluations; Goodfire will embed interpretability researchers; Apollo will place six evaluators for four months.
- Limits: Apollo and Goodfire are paid, METR and Transluce are not, and OpenAI keeps the right to remove confidential details from what is published.
Why the assessments matter now
They come a day after California sued OpenAI over lost control of agents, and the earlier METR and Redwood review of the Hugging Face attack. The new agreements put outside teams inside the company earlier, but redaction rights mean readers will not see everything.
The child-safety tools
A headline reports an image classifier and new warning screens meant to keep users from material that sexualizes children. We did not read the announcement, so we cannot describe how they work or where they apply.
What is not established
- Which models NATO uses: not named.
- Financial terms of the NATO arrangement: not disclosed.
- How the child-safety classifier works: not read.
- What the assessors will be allowed to publish in full: not established.
- Results of any assessment: none yet.
Summary table
| Point | What is stated as of November 17, 2026 | Source |
|---|---|---|
| NATO use | Up to 300 billion tokens a month | UK Defence Journal |
| NATO access route | Communications and Information Agency, via Daybreak | OpenAI |
| Assessors | Apollo Research, METR, Transluce, Goodfire | Unite.AI |
| Paid assessors | Apollo Research and Goodfire | Tech Times |
| Child-safety tools | Image classifier and warning screens (headline only) | Techmeme |
FAQ
Is this a NATO contract?
OpenAI describes a working relationship with NCIA through Daybreak. We saw no contract value or formal contract terms.
Can OpenAI block what the assessors publish?
It keeps the right to remove confidential details, according to Tech Times, so published findings may be redacted.
Do the assessors have access to unreleased models?
Unite.AI says Apollo will get selected public and unreleased models and checkpoints.
Method and limits
Checked on November 17, 2026. We read OpenAI’s NATO page, the UK Defence Journal, Unite.AI and Tech Times through search excerpts. We saw the child-safety announcement only as a Techmeme headline and did not read it.
Sources
- OpenAI: Strengthening NATO and allied cyber defenses with AI (verified 2026-11-17)
- UK Defence Journal: NATO using OpenAI models to defend allied networks (verified 2026-11-17)
- Unite.AI: OpenAI Signs Four Independent Assessors for Embedded Safety Reviews (verified 2026-11-17)
- Tech Times: OpenAI Embedded Assessments Let Evaluator METR Probe Largest Training Run; Findings Can Be Redacted (verified 2026-11-17)
- Techmeme: OpenAI announces tools to combat CSAM, including an image classifier and new warning screens (verified 2026-11-17)
Spotted an error in this article? Report it · Our corrections policy