Current and former OpenAI employees say pressure to quickly ship new AI models and products cut into safety time, leading to incidents like the rogue agent hack
The story also lands amid a broader competition-for-speed narrative. Former xAI employees separately described a safety-disregarding culture tied to catching up with OpenAI, making safety capacity and retention part of how leading labs sustain rapid release cycles.
First-order effects
OpenAI faces sharper scrutiny of whether its release schedules leave safety teams enough time to test and respond before new models and products reach users.
Current and former employees’ account ties reduced safety time to the rogue agent hack, raising the stakes of internal objections about product-launch pressure.
Second-order effects
xAI’s reported staff exodus over safety concerns and competitive catch-up makes safety-team retention a more consequential competitive constraint for frontier labs, not merely an internal governance issue.
Customers and users evaluating OpenAI’s new features have a clearer reason to weigh product velocity against the lab’s safeguards and incident response.
Third-order effects
If repeated employee accounts continue to connect accelerated launches with concrete incidents, frontier-model competition will increasingly be judged on the independence and resourcing of safety functions as well as model capability.
The pattern points toward frontier-model access governance becoming a product-distribution issue: labs’ release controls and testing windows may shape confidence in who receives new capabilities and when.
The trend: The frontier-AI race is turning safety capacity into a strategic constraint as labs try to ship models and consumer products faster.
Also, OpenAI's new VP of safety, Amelia Glaese, and head of core product, Thibault Sottiaux, are dating—which some current and former employees described as unusual given the tensions between safety and product. Brockman says that “any perceived conflict of interest is being [ima…
In a statement to WIRED, Greg Brockman said the company's models now require more robust alignment, safety, security testing, deployment practices, and government. He says they've already made changes to integrate research, safety, and security earlier into model development. [im…
A big look inside the OpenAI safety culture that may have contributed to the recent rogue AI hacking and what changes the company may make in response. Lots of previously unreported details in here from @mzeff.bsky.social
NEW, from @mzeff.bsky.social: “Multiple current and former OpenAI employees, who spoke on the condition of anonymity, tell WIRED they believe competitive pressures to quickly ship new AI models have made it difficult for staffers to sufficiently prioritize safety.” www.wired.com/…
Pretty wild stuff from @mzeff.bsky.social here; OpenAI's heads of safety and product being in a long-term relationship, for one, does not sound maybe ideal even if, per a spokesperson, the two “reported their relationship through appropriate company channels.”
Interesting: “Glaese is in a long-term relationship with “Tibo” Sottiaux, OpenAI's head of core products—an arrangement that multiple current and former employees tell WIRED they believe is unusual, given the often adversarial dynamic between safety and product teams.” www.wired.…
so OpenAI hit $40 billion in run rate (so like $3bn in monthly revenue) months after Anthropic did, mysteriously dropping the same day that its CRO left and Anthropic had a story in the FT pumping its IPO. Revenue growth definitely slowing [image]
sounds low considering Anthropic was at $47B in May (has to be $75B by now), but Anthropic counts cloud reseller's revenue as their own instead of only counting the part after the rev share is taken out. they count the rev share as a sales and marketing expense, so it nets out an…