OpenAI spells out the access it gives outside assessors of its safety cases

Visible chain of thought, together with internal deployment access for incident response and monitor red teaming, lets an outside team check whether a monitor flags the behaviour it is claimed to flag ^[2]. No volume of API querying produces that view. OpenAI says the access should let assessors challenge its assumptions, identify risks it may have missed, and reach their own conclusions about how well the safeguards work ^[3].

OpenAI sets the questions as well as the access. It says assessments are most useful when they address specific, consequential questions, and names three: whether the evidence supports a lab's safety case and claims, whether evaluations adequately test the risks they are intended to measure, and whether safeguards work under realistic conditions ^[4][5]. Those are the right questions to ask of a safeguard, and they are also the subject's framing of what counts as a finding.

OpenAI writes that labs have a responsibility to enable meaningful scrutiny while protecting sensitive information, and that labs and independent assessors share the responsibility for getting this right ^[10]. The post does not say who decides where that boundary falls, or which organisations have held the access so far ^[13].

The definitions are the testable part: a safety claim, as the post defines it, should identify the risks and conditions it addresses along with its assumptions and limitations ^[6]. A safety case connects individual claims to evidence and makes explicit the assumptions, uncertainties and remaining risks that could affect its conclusions ^[7].

What these engagements can do depends on their timing. Some run weeks and others several months, multiple in parallel, and the work is described as launch-agnostic, aimed at particular safety claims in depth over time ^[8][9]. The post allows that third party assessments can also be part of pre-deployment work and may inform deployment decisions ^[8]. That is a separate track from the one it sets out here. A finding from a launch-agnostic study arrives when the study ends, and the deployment it bears on may have shipped already.

The scope is private and non-profit assessment organisations doing technical safety work. OpenAI says its work with governments on testing and evaluation may call for different approaches because the roles and responsibilities differ ^[11]. Third party assessments are already part of the company's Preparedness Framework practices, and it says it has supported organisations and legislation pushing for a more rigorous and accountable process ^[14]. The first of the four priority areas is independent assessment of safety cases across training, evaluation, internal deployment and external deployment. It asks for expertise in alignment, control methods such as monitoring, cybersecurity, biological and chemical misuse, and red teaming, with multiple assessors likely examining different parts ^[12]. OpenAI says labs and assessors should be operating with shared international standards for safety and security practices ^[10].