The company racing to build superintelligence just fired its team responsible for making sure superintelligence doesn't kill us all.
The Summary
- OpenAI disbanded its preparedness team focused on AI safety and risk management as part of broader restructuring ahead of an expected IPO
- OpenAI and Anthropic are simultaneously limiting access to their most powerful models under US regulatory pressure
- The dual moves suggest OpenAI is optimizing for investor optics while downgrading internal safety infrastructure — exactly backward from what you'd want before releasing AGI into public markets
The Signal
OpenAI's preparedness team was created specifically to assess catastrophic risks from advanced AI systems. Not user harm or misinformation, but the genuinely scary stuff: rogue agents, runaway self-improvement, systems that develop goals misaligned with human survival. Now it's gone. The timing matters. OpenAI is restructuring ahead of what would be the most scrutinized tech IPO since Facebook.
The move comes as both OpenAI and Anthropic restrict access to their frontier models under pressure from US regulators. On the surface, access restrictions look like responsibility. In practice, they create a two-tier system: well-funded enterprises get the powerful models, independent researchers and startups get locked out.
"Access restrictions may hinder innovation and growth, affecting competitive dynamics and valuation in the AI sector."
Here's what's actually happening. OpenAI needs to show revenue discipline and centralized control to satisfy public market investors. Safety teams are cost centers. Distributed API access to researchers is a liability. The IPO roadshow demands clean narratives about monetization and moats, not messy realities about existential risk.
But the safety dilution raises fundamental questions about risk management when capital markets become the primary stakeholder. Public companies optimize for quarterly earnings. AGI labs need to optimize for not accidentally ending the world. These incentives don't align. At all.
The combined effect of disbanding internal safety review while restricting external scrutiny means fewer eyes on the most consequential technology humans have ever built. Independent researchers who might catch problems early lose model access. Internal teams who might flag risks before deployment lose organizational authority.
The Implication
Watch who OpenAI hires to replace the preparedness function. If it's a communications VP or compliance lawyer instead of an AI safety researcher, you'll know the restructuring was about narrative control, not capability improvement. The Web4 economy depends on powerful models being widely accessible to builders. If frontier AI access becomes a privilege reserved for enterprises that can afford six-figure contracts, the agent economy becomes a landlord economy. We've seen this movie before with Web2 platform APIs.
For anyone building agent infrastructure: assume you're designing for a world where the most capable models are either unavailable or untrustworthy. The companies that solve for local deployment, model distillation, and multi-model orchestration will own the next layer.