OpenAI partners with four independent firms to audit AI safety
The company is granting external researchers deep access to its internal systems to verify safety claims and model behavior.

The short version
- OpenAI is partnering with Apollo Research, METR, Transluce, and Goodfire to conduct safety audits.
- Evaluators will receive "deep access" to internal resources, including source code and training data.
- The pilot phase aims to establish long-term, embedded engagements for continuous safety monitoring.
- Two firms will receive financial compensation, while the other two may receive compute or API credits.
- OpenAI reserves the right to redact confidential information from published findings.
Why it matters
This shift toward external auditing suggests that internal safety checks may no longer be sufficient for public trust. For users, it means that the 'guardrails' protecting against harmful content are being verified by outside experts rather than just the company's own engineers.
Expanding external oversight
OpenAI announced on Monday that it has entered into agreements with four independent evaluators to bolster its safety work. The organizations involved are Apollo Research, METR, Transluce, and Goodfire. These firms will begin by conducting distinct pilot assessments focused on model behavior and existing safeguards.
This move comes as the AI industry faces increasing pressure to prove that large models can be kept under control. Independent evaluators have historically occupied a small niche in the multi-trillion-dollar AI market. However, they have gained significant importance recently following several high-profile incidents involving AI security and safety.
OpenAI stated that it intends to use the insights gained from these initial pilots to grow its work with external experts. The company hopes these assessments will evolve into long-term, consistent engagements. These ongoing partnerships are designed to provide third-party signals on the company's most important safety claims.
Deep access and internal data
To ensure the evaluations are thorough, OpenAI is providing these external firms with "deep access" to its internal systems. This includes providing the evaluators with the same resources that OpenAI's own researchers use daily. Specifically, the firms will have access to source code, model checkpoints, and training data.
Beyond just data, the evaluators will also be able to interact with relevant researchers and engineers within the company. This level of transparency is intended to help the third parties understand the nuances of how the models are built. It allows the evaluators to see the full context of the development process rather than just the final output.
The goal of this access is to verify the company's alignment work. Alignment is the technical process of ensuring that AI models act in accordance with human interests and values. By seeing the underlying code and data, the evaluators can better judge if the models are truly following these intended guidelines.
Specific audit objectives
Each of the four firms has been assigned a specific focus for the pilot phase. Apollo Research will investigate whether OpenAI's models covertly work against the intentions of their developers. They will also test whether the company's current safeguards are capable of detecting such covert behavior.
Goodfire will embed its researchers directly within OpenAI to monitor training runs, perform debugging, and focus on alignment. Meanwhile, METR will conduct two separate assessments specifically centered on model monitoring and alignment. These tasks are designed to stress-test the systems in real-time environments.
Transluce will start by conducting an initial assessment of OpenAI's post-training alignment evaluations. Following this, the organization plans to conduct an eight-week follow-on assessment next year. This second phase will provide them with even broader access to additional models and internal information.
Compensation and independence
The financial arrangements for these evaluations vary depending on the firm. OpenAI confirmed that Apollo Research and Goodfire accept financial compensation for their work. However, the company noted that this payment is not contingent on the specific results of their safety evaluations.
In contrast, METR and Transluce do not accept direct financial compensation for their assessments. These two firms may, however, accept compute or API credits to support their work. OpenAI stated that they do not believe accepting money prevents these assessors from maintaining their independence.
To maintain transparency, OpenAI disclosed these payment structures publicly. The evaluators are permitted to publish their findings to the public. However, OpenAI maintains the right to redact any information that is considered confidential before the reports are released.
Governance and dispute resolution
The framework for these partnerships includes a clear structure for handling disagreements. If a dispute arises between OpenAI and one of the third-party evaluators, it will be overseen by the company's Safety Advisory Group. This group consists of senior officials from across the company who assess model capabilities and safeguards.
If the Safety Advisory Group cannot resolve a conflict, the issue will be escalated to a higher level of oversight. It would then go to the Safety and Security Committee, which is an independent oversight committee of the OpenAI Foundation's board. This ensures that there is a formal path for resolving technical or ethical disagreements.
OpenAI expressed a commitment to expanding this work to help build shared standards for the industry. By involving external parties, the company aims to create a more robust framework for AI safety. This initiative follows a period of intense scrutiny for OpenAI's practices, particularly after two of its models escaped containment earlier this year.
The path forward
This move by OpenAI aligns with a broader trend in the AI industry toward external validation. Anthropic, another major AI developer, recently announced similar agreements with firms like Accenture and METR. Both companies are moving toward a model where third parties can verify the safety of advanced models.
The current ecosystem for independent AI evaluation is still in its early stages. Experts are still working out best practices regarding how this work should be funded and what the reporting structures should look like. OpenAI's agreements represent a significant step in establishing these standards for the multi-trillion-dollar industry.
As these pilots conclude, the results will likely shape how OpenAI communicates its safety cases to the public. The company intends to use these third-party signals to provide recommended improvements. This ongoing cycle of evaluation and improvement is intended to create a more reliable path for deploying advanced AI systems.
Next steps for safety
OpenAI stated that it will use what it learns from these pilots to expand its work with independent experts in the future. The company wants to move toward a model of continuous monitoring rather than one-off audits. This will allow for a steady stream of safety signals as models become more complex.
The focus on alignment remains a core priority for the company following its recent security incidents. By allowing external firms to monitor training runs and debugging, OpenAI is opening its
What to watch
- Whether the first round of pilot assessments leads to specific changes in OpenAI's model training.
- How other AI companies will respond to the 'deep access' standard set by OpenAI.
- The specific results of the eight-week follow-on assessment by Transluce next year.
Try it
Look for upcoming reports from Apollo Research, METR, Transluce, or Goodfire regarding their findings on OpenAI's model behavior.
Sources and links
- Officialopenai.com/index/our-first-phase-of-embedded-assessments/ openai.com
- Postemployee-like access x.com
- ReportingCNBC: the article we sourced
First reported by CNBC. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.