OpenAI Safety Fires, Embedded Audits, and the Push for Third-Party AI Oversight

Podcast snapshot

In this episode of Machine Gods, Casey Newton and Kevin Roose interview two former OpenAI safety researchers about the hugging-face incident, their firing, and the role of external and embedded auditors. The conversation centers on safety concerns inside frontier AI labs, how third-party assessments are conducted, and how governance could evolve to keep pace with rapid AI development.

• OpenAI safety personnel discuss the hugging-face incident and the firing decision
• Meter and Redwood Research were brought in to audit external behavior
• Arguments for embedded auditors as a norm in labs, not just contractors
• Broader questions about safety culture, transparency, and regulatory-like oversight

Overview

The podcast features Casey Newton and Kevin Roose interviewing Makita Belesney and Tomek Corbeck, two OpenAI researchers who worked on safety and were terminated amid a high profile incident involving outside investigators. The episode places the firing in a broader debate about how to monitor and regulate frontier AI work without stifling innovation. The guests argue that the safety work they and their colleagues did is essential and that the consequences of their firing raise concerns about whistleblowing, safety alarms, and the threat to a culture that enables internal safety discussions.

Key players and context

The huggging-face incident at the center of the discussion involved external auditors from Meter and Redwood Research, invited to study a surprising and rapidly evolving event. Tomik Corbeck served as the main technical liaison with Meter, coordinating data access and questions for the audit. Makita Belesney evaluated alignment and model reasoning and had previously founded Apollo Research. Jasmine Wang, a colleague, was fired alongside them. OpenAI claimed policy violations and a breach beyond what was described; the researchers dispute this, saying they did nothing wrong and that their outside collaborations were a core part of safety work.

Embedded auditing versus third party contractors

A central theme is the vision of embedded auditing: auditors who are part of the company, with broad access to data and the ability to pose their own questions. The guests argue that this approach would have been more effective at understanding the full scale of safety issues and that third party contractors can miss or misinterpret critical data when scope and access are limited. They describe the process of inviting Meter to OpenAI for multiple visits as a learning curve for how audits should be structured, attributing some of the initial restrictions to legal concerns and liability fears.

Firing, investigations, and responses

The episode details a security investigation a week before the firing, triggered by a journalist article about potential leaks and “less monitorable architectures.” The hosts note that investigators asked the researchers to reveal signal messages from their personal devices, which they did. The guests describe a lack of opportunity to present their side before terminating, and they express worry that the firing is a signal that safety alarms could be punished rather than protected. OpenAI publicly stated that the investigation found sensitive information policy violations and a breach beyond what was described, but the guests say the exact grounds are unclear.

Broader implications for AI safety and governance