OpenAI Dissolves Its Preparedness Team to Embed Safety into Product Labs
OpenAI reassigned its catastrophic risk researchers to core engineering groups.
OpenAI just dismantled its standalone Preparedness team, the specialized group that evaluated catastrophic risks from frontier AI models. Its leadership folded those biosecurity and cyber-risk researchers directly into everyday engineering units- instead of running safety checks from an isolated research island.
Critics naturally view any safety shuffle with skepticism, especially as OpenAI prepares for a public listing. Breaking up a dedicated watchdog unit sounds alarming on paper. Yet embedding risk researchers directly alongside model developers solves a major bottleneck in tech development.
Isolated ethics and safety teams acted as external auditors for years. They tested models only after engineering teams finished building them. That setup created friction and delayed software releases. By placing risk specialists directly inside core research teams, engineers spot vulnerabilities while writing code, rather than catching flaws right before launch.
Former Preparedness lead Dylan Scandinaro now focuses specifically on risks from self-improving AI systems. Meanwhile, specialized groups handle biosecurity and network defense directly within active model training pipelines. OpenAI president Greg Brockman argued that this tighter integration creates stronger safeguards across every new release.
Scattering safety talent across an enterprise carries real execution risks if leadership ignores internal warnings.
However, moving risk evaluation out of a separate silo and into daily development creates a hands-on security culture. AI safety works best when developers treat it as core product code rather than an afterthought.
