The NASK MoU Has No Audit Trail — and Europe's AI Rules Will Be Built On It
On a quiet Tuesday in the European policy cycle, Crypto Briefing published a story announcing that OpenAI and NASK — Poland's national research institute — had agreed to cooperate on AI safety research. I read it twice. Then I pulled the text into the same pipeline I use for protocol announcements and started marking fields.
There were two hard facts. One: the partnership exists. Two: NASK is a Polish institution. Everything else — scope, timeline, funding, research areas, whether NASK gets model access, which officials signed — was absent. The rest of the item read "may raise global AI safety standards," "may influence international AI policy." Two maybes in one paragraph.
In my line of work, an announcement with a two-field record is not a data point. It is a telemetry gap. I have audited token sales where a single missing modifier in a Solidity function meant the difference between a $5 million raise and a reentrancy drain. The lesson from 2017 never left me: the absence of information in a financial document is itself information. The only question is whether the absence is negligent or designed.
This is not a story about AI safety. It is a story about how a document with almost no verifiable content gets priced as a governance milestone — and why the crypto industry, of all audiences, should be the hardest crowd in the world to sell that to.
Context: a template, replicated
Some background, because the blanks only make sense against the playbook.
OpenAI has spent two years building what I would call safety diplomacy: bilateral, institution-to-institution arrangements that put frontier models in front of national evaluators before public release. The US AI Safety Institute and the UK AI Safety Institute both reached pre-deployment access arrangements in 2024. The Seoul frontier AI safety commitments added a layer of voluntary pledges. Each was a template. Templates get replicated, and replication is cheap.
NASK fits the mold of a partner but not the mold of a lab. Its institutional strengths sit in cybersecurity through CERT Polska, in disinformation research, and in Polish-language corpus work connected to PLLuM. That is applied security and social-impact expertise — genuinely scarce for low-resource languages and genuinely useful for red-teaming multilingual models. It is not the profile of an organization that studies model internals, training dynamics, or inference optimization. Which tells you, before any disclosure arrives, roughly where the work must sit: evaluation, red-teaming, alignment, watermarking and provenance. Not foundation model research.
So why does a crypto data desk care? Because the AI-crypto convergence trade is now a real market with real claims. I spent part of 2026 standardizing a benchmark dataset of 5,000 decentralized AI training jobs, and the single largest obstacle was not compute. It was that nobody could agree on what counted as evidence. Decentralized compute networks were publishing quality claims no independent party could reproduce, and every team measured variance its own way. The convergence sector is bottlenecked by verification, not by hardware.
Which makes an unverified AI safety partnership a strange thing to publish on a crypto outlet — and a useful specimen. It exhibits exactly the failure mode we have spent years auditing in DeFi: a promise with no attestation, marketed as infrastructure. And it arrived through a crypto news channel with no independent reporting, no interview, and no Web3 content whatsoever. That is the signature of a press release routed through a distribution list, not a reported story. I treat it accordingly.
Core: scoring the announcement
I treated the NASK item as a data event and scored it against the template established by previous OpenAI safety agreements. Five binary fields:
- Pre-deployment model access disclosed?
- Publication rights for research outputs disclosed?
- Exclusivity terms disclosed?
- Funding, compute, or in-kind consideration disclosed?
- Term and termination conditions disclosed?
The NASK announcement scores zero on all five. The US and UK arrangements score on at least two — model access and evaluation scope — even in their public summaries. Expressed as a query: