OpenAI is now letting outside researchers test early, less restricted versions of its models before they are actually released

[ControlPlane Leopard]

OpenAI has announced it will give independent researchers, specialist labs and domain experts controlled access to early model checkpoints before public release, including versions carrying fewer of the safety guardrails present in the final shipped product, allowing external testers to run their own evaluations that can now directly influence the company's eventual deployment decisions rather than simply reviewing an already finished model after the fact.

The new approach is structured around three distinct evaluation tracks. Independent lab evaluations allow research organisations to run their own custom tests specifically around autonomy, decision making and high risk domains such as cyber operations. A separate methodology review process has specialised reviewers examine the actual experimental setups behind larger scale studies and adversarial research. A third track brings in domain expert scoring, where professionals from fields including biosafety, medicine and cybersecurity assess whether a given model meaningfully elevates the capabilities of a relative novice attempting a specialised, potentially dangerous task.

OpenAI framed the change around the idea that AI capabilities are accelerating quickly enough that internal testing alone can no longer be relied on to surface every risk or blind spot on its own, arguing that independent evaluators bring fresh thinking, different incentives and genuine hands on experience with exactly the kind of high risk domains internal teams may not fully appreciate. The move comes shortly after public criticism from AI safety researchers questioning whether embedded evaluator arrangements at major labs, including OpenAI and Anthropic, can ever be considered genuinely independent given the financial and structural relationships involved.

Openai To Let Outside Groups Evaluate Ai Models At Earlier Phasewww.bloomberg.com

[Jackson79]

Giving outside evaluators access to versions with fewer safety guardrails specifically is the detail that actually matters here, that is a meaningfully different and more useful kind of access than just testing the fully polished final product.

[TheLegendBrett88]

The timing right after public criticism about whether embedded evaluators can genuinely stay independent feels less like coincidence and more like a fairly direct response to that exact pressure.

[Pilgrim11]

This feels like a genuinely positive step regardless of the underlying motivation behind it, external eyes on a system before release, even an imperfect version of that process, is better than the alternative of purely internal review.

[William]

Domain expert scoring around biosafety and cybersecurity specifically sounds like the genuinely highest value evaluation track here, that is exactly the kind of capability uplift risk internal red teams alone are least equipped to properly judge.

[Messi]

Independent evaluators bringing different incentives is a nice sentiment, but the actual financial relationship between OpenAI and whichever labs it chooses to grant this access to is exactly the kind of detail that determines whether that independence is real or just words

[Transformer Curtis]

Whether these external assessments can actually block or meaningfully delay a release, rather than just being noted and then overridden by commercial pressure, is the real test of how genuine this whole process actually is