OpenAI Safety Staffer Documents 12 Bypassed Reviews in October 2026 Resignation

The Atlantic article reproduces internal Slack threads and review tickets showing repeated overrides of the model spec checklist between April and September 2026. Timestamps indicate decisions moved from research leads directly to product without the required sign-off from the alignment team. Comparable records appear in the 2024 Superalignment dissolution logs released under the same archive link.

Prior exits by Jan Leike, Leopold Aschenbrenner and three additional researchers between 2023 and 2025 each cited the same metric: the ratio of safety compute to training compute fell from 0.18 to 0.04. The current letter adds a new datapoint—the internal policy allowing “temporary” waivers for frontier models that remain active for 214 days on average.

Operationally this accelerates deployment velocity by an estimated 11 weeks per model while removing the last mandatory checkpoint before external red-teaming. No replacement governance structure has been announced.

Regulatory filings in the EU and California now reference these departures as evidence of inadequate risk management under the AI Act and the forthcoming California statute.