Three OpenAI safety researchers — Tomek Korbak, Mikita Balesni, and Jasmine Wang — were fired over alleged leaks, which they deny and challenge as the official rationale for termination.

OpenAI says the terminations stem from violations of handling sensitive information and did not approve the use of Signal for employee communication, while not disclosing all contributing factors.

Two researchers defend talks about checking AI behavior and preserving readable reasoning, insisting they did not share confidential information; OpenAI maintains firings were due to broader sensitive-information policy violations, not just signaling or outside work.

The interview highlights ongoing debates about how third-party audits should operate and the risk of chilling collaboration with outside researchers.

AI safety groups and labs are formalizing external evaluations, with Anthropic agreeing to work with METR, signaling a broader industry shift toward third-party oversight.

The METR and outside safety network investigations underscore tensions between OpenAI and safety watchdogs as the field moves toward formal external model evaluations.

Listeners are directed to the Machine Gods transcript and platforms like YouTube, Spotify, and Apple for the full episode, with contact options for questions.

Researchers proposed continuous, embedded external oversight with broad access and confidential channels, arguing that limited reviews could deter staff from engaging with assessors.

NPR notes it does not offer or accept money for coverage or interviews and clarifies that such offers do not represent NPR.

The episode introduces Pandora’s Box and discusses wider Silicon Valley concerns, including Elon Musk’s SpaceX shares for American kids and new OpenAI bugging case filings involving chats with a Saudi royal.

Balesni indicated collaboration with Wang on cross-industry monitorability proposals and said he shared internal memos with OpenAI’s board and senior executives, reflecting their view of an oversight framework.

They warn the incident could set a precedent discouraging external investigations and ongoing auditing, citing OpenAI leaders’ discussions about embedded auditing and employee access.