What happened
NPR’s *Machine Gods* published an interview with Balesni and Korbak on October 23, 2026. NPR describes it as their first interview since their dismissal from OpenAI. The episode focuses on their account of events inside the company during METR’s investigation into a Hugging Face incident involving autonomous OpenAI agents, and on the broader disagreement over AI safety work at OpenAI.
NPR says the researchers explain why they believe OpenAI’s rebuttal does not add up. The episode frames the underlying dispute around the handling of sensitive information and the researchers’ interactions with METR. Balesni and Korbak say that their Signal contact with METR was part of their jobs, and they characterize the investigation as tightly scoped. NPR makes the episode audio available and directs listeners to a transcript.
Balesni and Korbak also make a broader argument about the form independent auditing should take. They say third-party auditors should operate more like continuously embedded employees than one-off contractors. The distinction is important: an embedded auditor may have more sustained exposure to systems and practices, while also raising questions about how independence is preserved when the audit function is situated within the organization it evaluates.
The interview follows OpenAI’s public explanation that the researchers were dismissed for violating policies on sensitive information. The researchers contest that account. NPR’s presentation records competing claims, not an independently established finding about the reasons for the dismissals or the full handling of the METR investigation. The episode runs 52 minutes and 50 seconds. NPR says listeners can access it through its own player and through YouTube, Spotify, and Apple Podcasts.
For builders, the operational issue is not whether every external review must use one organizational model. It is whether the review has a defined mandate, sufficient access, and a way to preserve its conclusions when those conclusions conflict with management. An audit limited to a narrow incident can be useful, but it cannot automatically support wider claims about a system’s safety practices.
What the archive adds
New Runtime’s earlier coverage on October 24, 2026, covered OpenAI’s defense of the dismissals and the researchers’ competing account. The NPR interview is a new source and a new account from two of the researchers, rather than a wholly new story. New Runtime also covered independent safety testing by Fangcun, a related example of external evaluation, though it concerns a different organization and model.
Trend signal
This adds a human and institutional dimension to the question of how safety claims are checked outside a model developer. The closest fit is New Runtime’s tracking of agent security at runtime boundaries: meaningful oversight depends on access, permissions, and operational boundaries, not just stated policy. The interview is a dispute about audit arrangements, not evidence that a particular runtime control works.
Why now
NPR presents this as Balesni and Korbak’s first interview after their dismissals, so it changes the public record from competing institutional statements to a direct account by two people involved. The timing also follows the public dispute over OpenAI’s explanation of the firings. It does not establish a final account of the events, but it makes the design of independent review a more visible practical question.
Connections to other events
New Runtime’s earlier coverage of Fangcun’s tests of DeepSeek V4.1 Flash-1017 described an external group finding failures involving shutdown software and monitoring gaps. That work and the METR investigation involve different systems, but both show why independent evaluation needs access to observable behavior and the ability to test controls rather than rely only on a developer’s assurances.
Why it matters
For teams building with AI, an audit is only as useful as its scope, access, and ability to report uncomfortable findings. A narrowly scoped investigation may answer a specific question without establishing broader safety; a continuous embedded arrangement may improve familiarity with systems but can complicate independence. Those are governance choices, not details to leave until after deployment.
The interview does not resolve which account of the dismissals is correct. It does, however, make clear that teams should define audit mandates in advance: what systems can be inspected, what incident evidence is retained, who controls communication with auditors, and how disagreements are documented. Teams should also decide who receives an auditor’s findings, whether the auditor can retain working records, and what can be published after a review. That makes later review more credible and helps separate a genuine security restriction from a constraint that could prevent meaningful scrutiny.
Limits and uncertainty
NPR’s page presents an interview and links to its audio and transcript; it does not independently adjudicate the researchers’ claims or OpenAI’s explanation. The account of Signal communications, the investigation’s exact boundaries, and the dismissal rationale remain contested. NPR does not quantify what the investigation found about the Hugging Face incident or establish how representative the researchers’ account is of OpenAI’s internal safety practices. Readers should treat the claims as attributed statements, not settled findings.
Skill to build now
Write a one-page audit charter for one AI workflow this week. Name the system and incident classes in scope; list the logs, prompts, tool calls, permissions, and evaluation artifacts an assessor may inspect; then specify who can restrict access and how that restriction is recorded. Add an escalation path for disagreements over findings. This is useful even without a formal outside auditor: it exposes whether a review can test the claims the team intends to make.
Reader takeaway
The story is less about a resolved safety finding than about who gets to inspect AI systems and under what conditions. For builders, the durable lesson is to specify auditor access, scope, record retention, and independence before an incident makes those terms part of a public dispute.
Open question
What evidence will establish the scope and findings of METR’s investigation and resolve the competing accounts of the researchers’ dismissals?