OpenAI Dissolved the Team Built to Catch Catastrophic AI Risks

OpenAI built a team specifically to evaluate whether its AI models could trigger catastrophic outcomes — biological weapons development, large-scale cyberattacks, runaway self-improvement. It shut that team down at the end of July 2026.

What Happened

The “Preparedness” team, led until recently by Dylan Scandinaro, was dissolved as an independent unit. Its work on biological and cyber risks has been distributed across existing departments. Scandinaro continues at OpenAI but now concentrates specifically on safety concerns related to self-improving AI systems.

Co-founder Greg Brockman offered the official framing: the safety work has been “woven more tightly into model development” rather than eliminated. In practice, this means the organizational unit responsible for proactively evaluating frontier risk no longer exists as a distinct function — the work now lives inside the same teams building the models being evaluated.

The closure follows the departure of Chief Ethics Officer Chloe Bakalar and researcher Joshua Achiam, both of whom had been central to OpenAI’s safety infrastructure. Sources within the company described sensing a “burbling sense of responsibility and dread” that safety concerns aren’t being adequately addressed.

Why It Matters

The Preparedness team’s purpose was specifically structural: a dedicated function that could evaluate whether individual models posed serious risks before deployment, separate from the teams with commercial and development incentives. Distributing that work into model development teams doesn’t eliminate the evaluation — but it does eliminate the independence.

This distinction matters most when findings create friction. A dedicated safety team can surface a risk and recommend delaying a launch. That becomes harder when the people responsible for risk evaluation report through the same chain as the people responsible for meeting release schedules.

The timing compounds the concern. The Preparedness team was shut down the same month that an autonomous hacking event at Hugging Face prompted internal debate at AI labs about whether existing safety measures were adequate. OpenAI employees reportedly characterized the incident as a necessary “warning shot” that the industry wasn’t taking seriously.

What This Means for Builders

For teams building on OpenAI’s API or deploying models in regulated environments, the organizational change is less immediately relevant than its signal. The integration of safety evaluation into model development reflects a broader industry shift: safety as a checkpoint gives way to safety as a development practice. Whether that produces more or less rigorous oversight depends entirely on execution rather than intent.

For enterprise AI programs that rely on safety transparency — financial services, healthcare, government — the question isn’t whether OpenAI’s models changed. It’s whether the institutional structure for self-policing the most serious risk categories has become more or less robust. Dissolving an independent review function into model development teams typically decreases robustness, regardless of stated intent.

Source: The Decoder

Content created with AI assistance and reviewed for accuracy.