OpenAI is opening a narrow window into its safety work. The company has signed agreements with four independent evaluators — among them NASK, a Polish state-supervised research institute — and announced a pilot program that CEO Sam Altman said would give outside researchers "employee-level" access to probe its artificial intelligence systems for risks.
The move lands as regulators in Brussels, Washington and London press frontier AI developers to submit to genuine third-party scrutiny, and as OpenAI's own safety commitments face renewed skepticism from researchers who say voluntary, company-managed access is not the same as independent evaluation.
What OpenAI Announced
According to reporting by CNBC, OpenAI has inked agreements with four independent evaluators to bolster its safety work, formalizing relationships that the company hopes will add outside rigor to its testing regime. Separately, Unite.AI reported that Poland's NASK — the National Research Institute formally known as Naukowa i Akademicka Sieć Komputerowa — has reached a deal with OpenAI on AI safety work, making it one of the named partners.
The centerpiece of the announcement, however, was Altman's pledge of "employee level" access: a commitment that outside researchers would be able to work with OpenAI's models and internal tooling at something approaching the privileges of a staff member. The details of how that access would be provisioned, which models would be covered, and what restrictions would apply were not fully specified.
The Skeptics' Case
NPR News reported that the pilot program "appears much more restrictive" than the framing suggested, and that researchers are already worried it won't be enough. That assessment captures the central tension: evaluations conducted at the invitation of the company being evaluated, under contractual terms the company sets, are structurally different from independent red-teaming with unmediated access to model weights, training data and deployment plans.
"The pilot program appears much more restrictive." — NPR News
Critics point to a familiar pattern in AI safety: announcements that emphasize access in principle while constraining it in practice. Common sticking points include non-disclosure agreements that prevent evaluators from publishing findings, access limited to selected models rather than frontier systems, and a review process that gives the developer veto power over what gets disclosed.
- Scope: Which models — and which capability thresholds — are actually in bounds for outside testing.
- Independence: Whether evaluators can publish negative findings without approval.
- Depth of access: Whether "employee level" includes weights, training pipelines and internal risk assessments, or only a sandboxed interface.
- Accountability: What happens if an evaluator finds a serious flaw, and who decides whether deployment proceeds.
How Different Outlets Framed It
The story's framing varied sharply by outlet. NPR led with the gap between announcement and substance, centering the anxieties of the research community and treating the program as a test of whether voluntary commitments mean anything. CNBC treated the four evaluator agreements as a business and governance milestone — evidence that OpenAI is institutionalizing safety partnerships as it scales. Unite.AI focused on the geographic and institutional dimension, highlighting NASK's role and the signal that a European, publicly funded institute is being brought inside a leading American lab's process.
None of the accounts described a program with binding authority. All four, read together, describe an arrangement that is more formal than informal advice but less than independent audit.
Why Poland's NASK Matters
NASK's involvement is significant beyond the bilateral deal. The institute has become a hub for European AI governance work, operating under Poland's digital affairs ministry and positioning itself within a widening network of national AI safety bodies — the UK's AI Security Institute, the US Center for AI Standards and Innovation, and counterparts across the EU — that are scrambling to build technical capacity to evaluate frontier models.
For OpenAI, partnering with a government-adjacent European institute offers credibility with EU regulators as the bloc's AI Act obligations for general-purpose models take effect. For NASK, the deal offers something scarcer: access to frontier systems that few public institutions can afford to train, let alone study.
The Larger Stakes
The context is a long-running dispute over whether AI labs can police themselves. In 2024, OpenAI's safety culture came under intense public scrutiny after high-profile departures from its safety teams, including researcher Jan Leike, who argued that safety had taken a back seat to product development. OpenAI has since restructured its preparedness framework and, in 2025, began describing its models as capable enough to warrant stronger safeguards at deployment.
Voluntary programs like this one are also a political instrument. As lawmakers weigh mandatory third-party auditing — a requirement several EU member states and US legislators have floated — industry-led access schemes can serve as a preemptive argument that regulation is unnecessary. That dynamic explains why safety researchers are pressing so hard on specifics: a program that looks robust on paper but limits what evaluators can see, say, or publish could inoculate OpenAI against tougher rules while changing little in practice.
What happens next will depend on details the announcement left open. If the four evaluators publish candid findings — including failures — the program could become a template for how frontier labs open their systems to outside scrutiny. If their reports stay private, the pilot risks becoming another entry in a longer record of safety commitments that sound stronger than they are.



