OpenAI disputes report that its centralized AI safety team was dissolved

[SašaJelenič]

A Financial Times report claiming OpenAI's centralized Preparedness team had been dissolved is being directly disputed by the company, which insists the team has not gone away, just been restructured, with research leaders covering cybersecurity, biology, chemistry and AI self improvement risk now reporting individually to Saachi Jain, OpenAI's head of safety, rather than sitting together inside one single consolidated group the way the team was originally structured when it was created back in 2023.

The Preparedness Framework itself is OpenAI's internal document defining specific capability thresholds across categories like cybersecurity, biological risk and persuasion, thresholds that are supposed to trigger specific additional safeguards, restricted access or, in the most extreme cases, an outright halt to release entirely once a model crosses into a genuinely dangerous capability tier. It is the exact same framework whose thresholds this week's separate Astra and Hugging Face incident announcements were explicitly measured against, which makes the timing of a dispute over the team's structural integrity considerably more consequential than it would otherwise be in a quieter week.

OpenAI's defense of the restructuring is that distributing specialized research leads directly under a single head of safety, rather than clustering them together inside one semi independent centralized team, actually strengthens rather than weakens institutional focus on catastrophic risk categories specifically, since it pulls those leads more directly and more visibly into core leadership decision making rather than leaving them somewhat siloed off to the side in a separate reporting structure. Whether that particular organizational argument holds up under real scrutiny probably depends heavily on details that remain genuinely unclear from the outside right now, specifically how much actual authority and independent budget those individual leads retain to flag serious concerns and meaningfully slow a release if they genuinely believe one is warranted.

The dispute is landing at a moment when OpenAI's public safety messaging is under about as much scrutiny as it has ever faced. In the very same week the company is disputing this specific organizational story, it is also publicly confirming a training pause tied to potential critical cybersecurity capability thresholds and simultaneously announcing new customer facing privacy and safety tooling in Private Safety Processing, three separate but clearly related storylines that all touch on how seriously and how visibly the company is treating its own stated safety commitments right now.

More than a thousand tech workers reportedly signed a petition in the aftermath calling for stronger external, government level oversight rather than continuing to rely primarily on each individual lab's own internal, voluntary structures and self reported commitments. That kind of organized internal and industry adjacent pushback, arriving in the very same week as a genuine capability threshold announcement and a specific organizational restructuring dispute, suggests plenty of people who actually work closely with or alongside this technology are not fully satisfied that internal structures alone, however they end up getting organized on an org chart, are sufficient on their own to catch problems reliably before they become serious enough to matter.

Openai Pause Astra Preparedness Frameworkwww.axios.com

[TheGame_Fan]

Reporting directly to the head of safety instead of sitting inside a semi independent team could genuinely go either way in practice depending entirely on execution, and honestly the framing alone does not tell you much either way. Sometimes flattening a hierarchy like that genuinely does increase visibility and speed of escalation, sometimes it just quietly dilutes independent voice and authority under the appearance of consolidation, hard to know from outside which one this actually is.