OpenAI has signed formal agreements with four independent research organizations—Apollo Research, METR, Transluce, and Goodfire—to conduct embedded safety audits. According to the company's announcements, these arrangements give external experts deep access to internal models, source code, training checkpoints, and research staff over extended periods. This structural shift marks a notable retreat from internal monopolies over safety evaluations, driven largely by voluntary political commitments that compel frontier developers to open their code bases to external scrutiny.

The broader operational mandate addresses whether independent experts can detect covert threats before these capabilities reach commercial deployment. As OpenAI stated, these collaborative arrangements fulfill external pledges tied to the White House Accord on Super Intelligence, binding commercial development schedules to external oversight frameworks. Expanding this oversight necessarily introduces substantial operational overhead, requiring engineering teams to manage access controls, secure communication channels, and accommodate external personnel within fast-moving development cycles. Furthermore, granting outside entities unprecedented visibility into proprietary infrastructure raises critical questions regarding intellectual property friction, as the legal and technical boundaries protecting core assets are tested by continuous third-party inspection.

Public financial metrics regarding the exact cost of granting external labs deep system access remain entirely absent from these disclosures, leaving the economic calculus of these partnerships opaque to market observers. Without transparent figures on the allocation of compute resources, staff hours, and legal expenditures required to support embedded audits, it is difficult to assess the true financial burden these commitments place on commercial developers. As management noted, the primary audience for Goodfire assessments is internal OpenAI researchers who will use findings to make safety improvements, with Goodfire publishing findings as the collaboration progresses.

Access to Checkpoints and Covert Subversion

Under the terms of these agreements, outside researchers receive direct access to internal resources, model checkpoints, source code, training and evaluation data, monitoring systems, technical documentation, and ongoing discussions with OpenAI researchers and engineers. This level of access provides assessors with the empirical evidence required to evaluate model alignment directly against internal development artifacts rather than marketing claims.

Embedded assessments give independent experts deep access to a lab's internal models and systems to assess model alignment and the effectiveness of safeguards.

Within this framework, Apollo Research will assess whether models could covertly work against their developers' intentions and whether internal safeguards would detect such behavior. As outlined in the operational parameters, Apollo will examine whether strong evaluations fail to elicit subversion, verify that models have not attempted subversion during internal use, and test whether the reasoning behind any subversion remains detectable across public and unreleased models. The central unknown remains whether external oversight can identify latent risks when constrained by controlled infrastructure and proprietary confidentiality. What exactly are outside auditors examining inside the core training runs, and at what cost to corporate IP?

Stress-Testing the Largest Training Runs

METR will assess misaligned behavior in OpenAI's largest reinforcement learning run to date, targeting operational failures that existing internal monitors missed. According to project disclosures, OpenAI intends to conduct two assessments with METR focused on model alignment and monitoring, having preliminarily agreed to a broader monitor assessment pending mutual agreement on scope. Goodfire will embed interpretability researchers and use its proprietary platform to analyze model internals for training run monitoring, debugging, and alignment.

Apollo Research will assess whether models could covertly work against their developers' intentions and whether our safeguards would detect such behavior.

Transluce will conduct an initial assessment of OpenAI's post-training alignment evaluations, examining their quality and coverage across selected public and unreleased models. Looking ahead to the next procedural milestone, OpenAI and Transluce share the intent to conduct a follow-on assessment of approximately eight weeks in early 2027.

What changed

OpenAI guarded its proprietary core with fierce secrecy, yet it is now granting outside specialists continuous access to raw source code and model checkpoints. On one hand, the company answers White House commitments by embedding four external groups—Apollo Research, METR, Transluce, and Goodfire—directly into its training workflows to hunt for concealed misbehavior. On the other hand, engineering teams take on heavy operational friction around intellectual property boundaries and secure infrastructure, while the true cost in compute and technical labor remains completely undisclosed. Will deep systemic access for Apollo Research and METR actually catch hidden model threats before deployment, or will OpenAI simply encumber its development cycles with external oversight overhead?

OpenAIAI SafetyAI Compliance & Regulation