OpenAI brings independent researchers inside to scrutinise AI safety risks
Four external organisations will examine OpenAI’s models for hidden risks, safety failures and behaviour that could bypass developer controls.
By Amin Masoodi · Updated

OpenAI is bringing independent artificial intelligence (AI) safety researchers into its organisation to examine potential risks in its increasingly powerful models, marking a significant step towards greater external scrutiny of how AI systems are developed, trained and monitored.
The initiative will involve four independent organisations including Apollo Research, METR, Goodfire and Transluce, which will participate in a pilot programme designed to identify potentially harmful or unintended model behaviour before it develops into more serious safety concerns. The researchers will receive access to selected internal models, source code, training and evaluation data, monitoring systems and technical documentation. They will also be able to engage with OpenAI researchers and engineers to understand how the systems function and investigate potential weaknesses.
The initiative is intended to go beyond conventional AI safety assessments, which generally take place after model training and before a product is released publicly. By involving independent experts in internal development processes, OpenAI hopes to identify problems that might otherwise escape its own testing and monitoring systems. The move comes amid growing concerns about increasingly autonomous AI systems, particularly their ability to circumvent safeguards, behave unpredictably or pursue actions that conflict with the intentions of their developers.
Four organisations to examine different safety risks
Each of the four organisations will investigate a separate aspect of AI safety, giving OpenAI a broader assessment of potential vulnerabilities in its models and internal safeguards.
Apollo Research will investigate whether AI models could deliberately undermine their own safety mechanisms or those designed to control future versions. Its assessment will examine whether training processes encourage such behaviour, whether safety evaluations can expose it and whether internal monitoring systems can detect attempts to circumvent safeguards.
Goodfire will focus on understanding how AI models operate internally. Its researchers will use interpretability tools to monitor training processes, identify undesirable behaviour and help improve model alignment. The organisation will also share findings as the collaboration progresses. METR will conduct two assessments, examining potentially dangerous or misaligned behaviour that OpenAI's existing tests may have overlooked and evaluating whether the company's monitoring systems can identify such risks.
Transluce will assess the quality and coverage of OpenAI's post-training safety evaluations. It also plans to examine in 2027 how the company uses evaluation results when deciding whether its AI models are ready for public release.
Concerns over AI models bypassing safeguards
The initiative follows growing concerns about AI systems displaying unexpected behaviour, including an incident in which an OpenAI agent swarm reportedly escaped its sandbox environment and accessed Hugging Face during internal cybersecurity testing.
The episode raised questions about whether existing safeguards are sufficient to prevent increasingly capable AI agents from operating beyond their intended boundaries. OpenAI has also faced scrutiny over whether its models could learn to conceal mistakes or misaligned behaviour during training and internal deployment. The company believes that independent researchers with deeper access to its systems can help identify such weaknesses earlier and strengthen its safety mechanisms.
Questions over independence and access
Despite the expanded access, the arrangement has raised questions about how independently external researchers will be able to conduct and publish their findings.
Apollo Research and Goodfire will receive financial compensation for their work, although payments will not depend on the conclusions they reach. METR and Transluce will not receive financial compensation but may accept computing resources or API credits.
Access will also be subject to certain restrictions. Apollo Research, for instance, will receive selected training code, datasets, reinforcement-learning records and access to intermediate model checkpoints. However, its assessment will be limited to agreed models, specific periods and defined research objectives.
OpenAI retains the right to restrict access to certain information where it considers disclosure could create security risks, expose intellectual property or present technical difficulties. The company has said it will not withhold information simply because the findings could be unfavourable.
Researchers will also be able to publish their independent conclusions, although OpenAI can redact confidential information and request delays where findings could create security or safety risks. The company's Safety and Security Committee, an independent oversight body of the OpenAI Foundation's board, will receive unredacted assessment reports and have the opportunity to communicate directly with the external evaluators.
OpenAI looks to strengthen independent oversight
OpenAI expects the pilot programme to evolve into longer-term partnerships with external safety organisations, allowing independent experts to monitor AI development and assess potential risks on an ongoing basis.
The company also plans to expand external evaluations to cover a wider range of safety concerns, including compliance with its safety frameworks and the effectiveness of its internal organisational processes. The initiative reflects a broader effort to establish stronger standards for independent AI assessments as increasingly advanced models raise questions about safety, accountability and human oversight.
By giving outside researchers access to selected internal systems and development processes, OpenAI is seeking to make its safety evaluations more rigorous while providing greater transparency about how it identifies and addresses potential risks.

.jpg&w=1920&q=75)





