When Safety Becomes the Cost of Going Public: The Deeper Signal Behind OpenAI Disbanding Its Preparedness Team

A Story That Shouldn't Be Buried

On August 19, 2026, The Next Web reported what should have been an industry-shaking piece of news—but was quickly drowned out by the daily noise: OpenAI has disbanded its Preparedness Team, the group specifically responsible for evaluating catastrophic model risks.

The reason? "IPO streamlining."

Six words. Nonchalant. Earth-shattering.

What Was the Preparedness Team?

To understand the weight of this decision, you first need to understand what this team did.

The Preparedness Team was established in late 2023 as an institutional response to OpenAI's internal safety turmoil. Its mandate was crystal clear: before any model release, evaluate its potential for catastrophic risk—including but not limited to biological weapons assistance, automated cyberattacks, mass disinformation generation, and loss-of-control scenarios with autonomous AI systems.

In simple terms, the team existed to answer one critical question: Could someone use this model to do something extremely dangerous?

This was not some optional "compliance decoration." In an era of exponential AI capability growth, catastrophic risk assessment is the last line of defense in the entire safety architecture. Disbanding this team is like removing the safety valve from a nuclear reactor—it may not explode in the short term, but if something goes wrong, the consequences are irreversible.

A Troubling Timeline

When we zoom out, a disturbing pattern emerges:

May 2024: OpenAI's head of safety Jan Leike and co-founder Ilya Sutskever both depart. Leike's resignation letter stated bluntly that "safety culture has taken a backseat to shiny products."

June 2024: Former employees publish an open letter warning that OpenAI is "prioritizing commercial interests over safety."

November 2025: At the Seoul AI Safety Summit, OpenAI and other major AI companies sign voluntary safety commitments, including establishing catastrophic risk assessment mechanisms.

June 2026: OpenAI announces IPO plans, with a valuation potentially exceeding $300 billion.

August 2026: The Preparedness Team is disbanded.

From safety leadership departures, to weakened safety commitments, to the outright elimination of safety teams—this is not an isolated incident, but a systematic trend. Each step has a "reasonable" business justification, but taken together, the direction is unmistakably clear.

The Invisible Pressure of Going Public

The key to understanding all of this lies in the IPO.

When a company prepares to go public, it faces multidimensional pressures: profitability, growth velocity, organizational efficiency, cost control. And a team whose primary job is to "halt releases" is inherently "inefficient" in IPO logic—it generates no revenue, yet can delay product launches and affect quarterly numbers.

This is the fundamental dilemma facing AI safety: The value of safety work manifests in "things that didn't happen," while capital markets only pay for "things that did."

A successful catastrophic risk assessment won't appear in a prospectus—"We discovered and patched a vulnerability that could have enabled bioweapons proliferation" isn't sexy; "Monthly active users surpassed 1 billion" is. But a failed risk assessment? That could be an irreversible catastrophe.

OpenAI's choice is entirely "rational" in business logic. But this rationality itself exposes a fundamental flaw in the current AI governance framework: we have entrusted humanity's most critical safety guardrail to market-driven entities whose primary obligation is maximizing profit.

Industry Ripple Effects

OpenAI's move will not exist in isolation. It will be interpreted and amplified across the entire industry.

For other AI companies: When the industry leader deems a catastrophic risk assessment team "streamlinable," what justification do other companies have for maintaining similar teams? This could trigger a wave of safety team layoffs—not because these teams are unimportant, but because "even OpenAI did it."

For regulators: This provides the most direct argument for regulatory intervention. If companies are unwilling to self-regulate, the necessity of external oversight becomes even clearer. The EU AI Act's enforcement may intensify, and the US Congress could accelerate AI safety legislation.

For the AI safety research community: A disheartening but potentially catalyzing signal. When internal corporate safety mechanisms are weakened, the importance of independent safety research and evaluation institutions rises correspondingly.

A Bigger Picture: Anthropic's Same-Day Troubles

The same day's AI news reveals another dimension of the industry's困境.

Anthropic—the company that has always marketed itself on "safety"—is simultaneously facing a triple crisis: Claude Code weekly usage limits reduced by a third, AI output watermarks triggering a user cancellation wave, and performance degradation across multiple Claude models.

This creates a bitter parallel with OpenAI's situation: Even companies renowned for safety are forced to make decisions that compromise user experience and trust under commercial pressure.

When safety teams can be disbanded, when user experience can be sacrificed, when watermarks can be forcibly added without user consent—we must ask a more fundamental question: Are AI companies' commitments to safety and users genuine values, or just marketing tools?

The Questions We Actually Need to Discuss

OpenAI's disbanding of the Preparedness Team is ostensibly an organizational adjustment, but it touches on several core propositions of AI governance:

First, should catastrophic risk assessment be a voluntary corporate action or a mandatory requirement? If the former, today's events prove the fragility of this model. Companies can abandon safety commitments at any time, for any reason, without substantive consequences.

Second, can IPOs and safety coexist? There is a structural contradiction between the short-term pressures of capital markets and the long-term requirements of AI safety. Current institutional design offers no solution to this contradiction.

Third, who guards the last line of defense? If internal safety teams can be disbanded, should independent third-party evaluation mechanisms become mandatory?

These questions have no simple answers, but they demand serious discussion—not quiet erasure amid the noise of an IPO.

Closing Thought

On August 19, 2026, OpenAI disbanded its catastrophic risk assessment team. Perhaps in retrospect, years from now, this will be seen as an insignificant organizational tweak. Or perhaps it will mark the beginning of something far larger.

AI capabilities are growing exponentially, while the team responsible for evaluating whether those capabilities could lead to catastrophic consequences is being "streamlined."

This contradiction will not disappear simply because we choose not to discuss it.

For everyone who cares about AI's future—developers, researchers, policymakers, ordinary users—today's news deserves serious attention. Not because it necessarily means catastrophe is imminent, but because it reveals a trend we must confront: When safety becomes friction for business, safety is always the first to be sacrificed.

And that, in itself, is the most dangerous signal of all.