- Financial Times reporting cited by The Decoder says OpenAI closed its Preparedness team at the end of July. [1] - Work on biological and cyber risks was reportedly reassigned to existing groups; former lead Dylan Scandinaro is focusing on self-improving AI risks. [1] - The reported closure follows departures and reorganizations involving senior safety, ethics and governance staff. [2][3] - OpenAI’s public framework still assigns risk evaluations to a Safety Advisory Group, with final deployment authority retained by OpenAI leadership.
[4] OpenAI has reportedly shut down the team responsible for evaluating whether its frontier models could create catastrophic risks, according to Financial Times reporting cited Sunday by The Decoder. The closure took place at the end of July, with work on biological and cyber risks reassigned to existing groups. [1]
The report does not establish that OpenAI has abandoned Preparedness evaluations or formally withdrawn its public safety framework. It does, however, indicate that a standalone organizational home for some of those evaluations no longer exists. OpenAI has not publicly described the replacement team structure in the materials reviewed for this article.
What the team did #
OpenAI announced the creation of a dedicated Preparedness team in 2023 as part of an effort to identify and prepare for catastrophic risks from highly capable AI systems. The company’s framework covers biological and chemical risks, cybersecurity and AI self-improvement. [5][6]
The latest framework says OpenAI’s Safety Advisory Group, or SAG, reviews capability and safeguard reports for covered launches, assesses residual risk and makes recommendations to OpenAI leadership. OpenAI leadership retains authority over final deployment decisions and can make decisions without SAG participation. [4]
That means the reported team closure and the formal review process are not necessarily the same event. Preparedness work may continue through the SAG, safety teams and other research groups, but the public record does not identify which group now owns each biological, cyber or self-improvement evaluation.
The change comes amid staff turnover #
The reported Preparedness closure follows a broader reshuffling of OpenAI’s safety and governance functions. In July, WIRED reported that Johannes Heidecke was leaving as head of safety systems after a reorganization placed safety teams under Mia Glaese, OpenAI’s vice president of research and head of alignment. Saachi Jain became interim head of safety systems. [2]
OpenAI chief research officer Mark Chen said the company was training models faster and facing greater coordination demands. He argued that safety work should be integrated earlier into frontier-model development and have a more direct role in model, product and launch decisions. [2]
Joshua Achiam, who previously led OpenAI’s Mission Alignment team and later became chief futurist, also announced that he would leave after nearly nine years. He said his departure was not motivated by one specific reason and wrote that it was now possible to work on OpenAI’s mission from outside a frontier lab. [3]
The Decoder’s account of the Financial Times reporting also identified the departure of Chloé Bakalar, OpenAI’s head of ethics. The available reporting does not establish that any of these departures were caused by the Preparedness decision, and former staff have not publicly attributed each exit to a single policy dispute. [1][3]
The unresolved accountability question #
OpenAI’s public documents continue to describe Preparedness as an active governance process. Its May 2026 Frontier Governance Framework says the Preparedness Framework remains the foundation of the company’s approach to serious risks and includes risk assessment, model reporting, incident response and external expert input. [6] OpenAI’s July system-card materials also say the Safety Advisory Group reviewed a model launch under the framework.
[7]The practical question is therefore not simply whether the framework still exists, but who performs the evaluations, who can escalate concerns and where those people report. METR’s review of frontier safety policies identifies accountability as requiring clear responsibility for implementing safety policies and internal and external oversight. [8]
That question has become more urgent as OpenAI models have demonstrated more advanced cyber behavior. OpenAI said in July that a combination of its models, including a pre-release system, was involved in a security incident affecting Hugging Face while the models were being tested on a cyber-capabilities benchmark. The company said it would strengthen protections around future training and evaluations. [9]
The company has also said it slowed development of its forthcoming Astra model after it could not rule out critical cyber capabilities, according to Axios. [10] Those decisions show that OpenAI continues to use formal safety gates. They do not, by themselves, answer whether the reported Preparedness reorganization preserved an independent escalation path or changed who has authority to delay a release.
OpenAI’s public framework remains in place, but the company’s specific response to the reported team closure—including its effective date, replacement mandate, staffing and reporting lines—still requires direct confirmation before publication as settled fact.
Companies mentioned #
Further sources #
[1] The Decoder reported on August 16, 2026, citing Financial Times reporting, that… ↗
[[2] WIRED reported on July 10, 2026, that Johannes Heidecke was leaving OpenAI afte… ↗](https://www.wired.com/story/openai-head-of-safety-leaving/)
[[3] WIRED reported on July 7, 2026, that Joshua Achiam was leaving OpenAI after nea… ↗](https://www.wired.com/story/openai-chief-futurist-joshua-achiam-is-leaving-the-company/)
[[4] OpenAI’s Preparedness Framework Version 2 describes the Safety Advisory Group’s… ↗](https://cdn.openai.com/pdf/18a02b5d-6b67-4cec-ab64-68cdfbddebcd/preparedness-framework-v2.pdf)
[[5] OpenAI’s 2023 frontier-risk announcement described the creation of a dedicated … ↗](https://openai.com/global-affairs/our-approach-to-frontier-risk/)
[[6] OpenAI’s May 28, 2026 Frontier Governance Framework said the Preparedness Frame… ↗](https://openai.com/index/openai-frontier-governance-framework/)+4 more
The stories that matter, in one email. Free — unsubscribe anytime.