Hook
Safety is the only truth the market respects. On Wednesday, OpenAI confirmed the dissolution of its Preparedness team—the unit tasked with identifying catastrophic risks from frontier models. The move comes months after the Superalignment team was gutted. Two safety teams dead in twelve months. The signal is unambiguous: the IPO clock is ticking, and safety governance is being restructured as a cost center.
Context
Founded in 2023, the Preparedness team was OpenAI's primary defense against bioweapons, cyberattacks, persuasion manipulation, and autonomous replication risks. It reported directly to the Safety and Security Committee, a board-level body. Its leader, Aleksander Madry, was quietly moved off core safety decisions in late 2024. Now the team is gone. OpenAI claims the functions are being "absorbed" into other departments. But absorption without a dedicated unit is a demotion. The timing is everything: a rumored IPO in 2026 demands leaner org charts, higher profit margins, and fewer unmonetized teams.
Core
Let me be explicit. The Preparedness team's budget was not trivial. Senior researchers at OpenAI command $500k+ total compensation. Red-teaming at scale requires GPU clusters. Both are pure cost with no direct revenue. In a pre-IPO restructuring, zero-revenue teams are the first to be cut. Quantitatively, the ratio of safety researchers to product engineers has dropped by an estimated 40% since January 2025. This is not a conspiracy theory; it's a line-item decision.
But the immediate impact is not just about cost. The Preparedness team was the gatekeeper for model releases. They ran the adversarial testing that prevented catastrophic jailbreaks. Without them, the release cadence accelerates—but so does the tail risk of a major safety incident. Based on my audit experience across crypto protocols, I've seen this pattern before: when a security team is disbanded, incident frequency spikes within six months. The same logic applies to AI. The market is pricing in a 15% probability of a significant model failure in the next twelve months, according to prediction markets. This move doubles that probability.
Contrarian
Here's the angle the mainstream coverage misses. The Preparedness team's dissolution is not purely a loss. It could accelerate the growth of a third-party AI safety audit industry. Just as crypto exchanges started outsourcing smart contract audits after the DAO hack, AI companies will now hire external red teams, model insurance firms, and compliance consultants. The faucet runs dry, the dryers crack. OpenAI's internal capacity is gone, but the demand for safety assurance remains. Companies like Anthropic, which still has a robust safety team, will use this as a competitive differentiator in enterprise RFPs. Financial institutions, healthcare providers, and government agencies will demand external validation. The safety market is being decoupled from the model market.
Moreover, the narrative that "safety is sacrificed for IPO" is too simplistic. The board may have calculated that the Preparedness team's methods were too conservative, slowing down product releases. By disbanding it, they are betting that the marginal safety risk is acceptable relative to the IPO timeline. That is a rational, if cynical, calculation. The real question is whether the market will punish them for it. In the short term, probably not. In the long term, a single catastrophic failure could erase the entire IPO value.
Takeaway
Watch for two things: first, the departure of key Preparedness team members to competitors or to independent safety startups. Second, the next major model release—if it comes without a detailed safety report, the narrative is confirmed. Safety is not being abandoned; it's being outsourced. But outsourcing oversight is the same as delegating trust. And trust, once lost, is not easily regained. Chasing ghosts in the digital art auction house is one thing. Chasing ghosts in the safety data is another. Leading the charge when the herd turns away means being the one who reads the risk report before the model is deployed.