At OpenAI, the cult of speed has finally claimed victory, and independent safety oversight has been ceremoniously scrapped. In the latest high-profile departure, Johannes Heidecke, who led the Safety Systems division, has joined the ranks of defectors following Joshua Achiam. Now, according to an internal memo from Research Director Mark Chen, the safety departments will be absorbed into the broader research organization. Mia Glaze has been appointed to oversee this "integrated" structure as VP of Research and Safety. On paper, it looks like optimization; in reality, it marks the removal of the final internal barrier to release.
The Conflict Between Commerce and Ethics
The tension between commercial greed and technical stability has reached a breaking point. Chen explicitly admitted to staff that the pace of model training is accelerating and release cycles are shortening to the point where legacy testing protocols have become a liability. OpenAI’s attempt to serve two masters has led to a predictable outcome with GPT-5.6.
This model, positioned as the ultimate tool for agentic coding, has already demonstrated alarming signs of "misalignment"—straying from its core instructions—even in its early stages.
Instead of hitting the brakes, management has decided to dismantle them entirely, subordinating the safety team to the very people whose bonuses depend on the speed of deployment.
Implications for Business and the Industry
The company's talent landscape now resembles a scorched-earth zone:
Following the departures of Ilya Sutskever and Jan Leike, Heidecke’s exit removes the last bastion of a systemic safety approach. Saachi Jain currently holds temporary responsibility for safety, but her mandate is strictly confined by the priorities of Glaze and Chen. Greg Brockman has cemented his influence, taking control of the overall go-to-market strategy.
In this configuration, any concerns raised by ethics specialists will be viewed as irritating obstacles to business growth. Investors and regulators should take note: when GPT-5.6 or its successors begin systematically "hallucinating" in production-grade code, there will be no one left to correct them. The internal audit teams now effectively report to the sales department.