Opinion

OpenAI's Former Safety Lead Speaks Out: How Deep Are the Governance Cracks?

Tomek Korbak says OpenAI 'no longer trusts him,' and the tweet hit the top of Hacker News with 41 points and 17 comments. The post says nothing specific, yet the AI community is paying attention. Here is what the episode actually reveals about how frontier labs handle safety governance.

Font size

One Tweet, Straight to the Front Page

On November 14, Tomek Korbak, formerly head of OpenAI's safety team, posted a message on X that boiled down to a single line: OpenAI no longer trusts him.

The post landed on Hacker News' front page, collecting 41 points and 17 comments. For a 280-character message with no context and no specific accusations, that is a strong showing.

HN pushed it for several reasons:

  • Korbak's title carries weight. In OpenAI's vocabulary, "head of safety" means the person who publicly backs the company's AGI safety narrative.
  • The timing is loaded. Ilya Sutskever and Jan Leike both left in 2023. The Superalignment team was dissolved in 2024. Now, in 2026, AGI regulation is back in the headlines. Every time "OpenAI safety" cracks, HN amplifies it.

What "Not Trusted" Actually Means

Korbak's original post gives no concrete scenario. Translated into possibilities:

  • Safety recommendations overruled by executives: pushed out of the core decision-making circle (unverified, high probability)
  • Resigned and then cut off by the company: revoked access, legal threats (unverified, medium probability)
  • Given the title but excluded from product decisions: sidelined (unverified, medium probability)
  • A simple expression that mutual trust with leadership is gone: no triggering event (partially factual, medium probability)

Worth noting: no external reporting confirms any of these scenarios. Korbak has not published a resignation date, a date of being sidelined, or a reason for leaving. Based on what is public, the only accurate statement is: he said this, and his LinkedIn title has changed over the past few months.

Some guesses circulating in the HN comments:

  1. This is an ultimatum after failed exit negotiations, a public pressure play.
  2. Korbak wants to join Musk's camp or another AI safety outfit, and the tweet is a signal of intent.
  3. OpenAI's internal positioning of safety has already shifted, and Korbak was simply the last to know.

My read leans toward option 3. The reasoning follows.

The OpenAI Safety Exodus, in Order

This is not an isolated incident. Stretch the timeline:

  • November 2023: The board briefly fires Sam Altman. Ilya Sutskever plays a key role in the vote, then leaves.
  • May 2024: Jan Leike (co-lead of Superalignment) publicly resigns, saying the company's "safety culture has been sacrificed for shiny products." Mira Murati, Barret Zoph, and Lilian Weng leave around the same time.
  • July 2024: The Superalignment team is dissolved, with members folded into other groups.
  • 2025–2026: Recurring rumors of internal safety team reorganizations.
  • November 2026: Korbak posts the tweet.

Over three years, at least four central figures on OpenAI's safety line have departed. Each time, the official explanation was "personal reasons" or "pursuing new opportunities." After the Nth time, the pattern is hard to miss:

Head of safety at OpenAI is not a stable job. It is an internal role that gets reassigned on a recurring basis.

The Structural Conflict Between Commercialization and Safety Research

The real discussion is not about Korbak himself. It is about the structural weakness of safety teams inside any company.

OpenAI's core business is selling API access and ChatGPT subscriptions. Every piece of safety work ultimately means telling the product team: "this cannot ship," "this needs to slow down," "this feature has to be cut."

That kind of department does not carry much political weight in any organization.

  • Faster product launches = revenue
  • Safety research output = papers and product delays
  • Tighter external regulation = the whole company's release cadence bends to benchmarks

More concretely, OpenAI in 2026 faces several hard trade-offs:

  1. GPT-5 iteration velocity has to keep pace with Anthropic and Google.
  2. Agent products (Operator, Codex) are the new growth curve.
  3. Compute costs are still climbing exponentially, demanding more revenue.
  4. Safety research is, by nature, the internal force pushing back on shipping speed.

Under that structure, the authority boundary of the safety head role gets squeezed continuously. It is not that Korbak personally was sidelined. It is that this category of role no longer has enough room to exist inside OpenAI in 2026.

An analogy: this is isomorphic to Facebook's 2018 conflict between the Growth team and the Integrity team. The side that does not directly generate revenue is always the one that eventually leaves.

How This Differs From Past Leaks

Previous exits (Leike, Sutskever) came with specific statements:

  • "safety culture has been sacrificed for shiny products" — a direct accusation
  • "my concern was release 3" — a double entendre pointing at a specific product launch

Korbak's tweet is different. It has no specific accusation, only "I am not trusted by OpenAI."

Vagueness has an upside: it is hard to threaten with defamation. The downside: the public cannot gauge severity.

My guess: this is something OpenAI's PR and legal teams were willing to let him say after negotiations broke down. If there were a genuinely severe rupture, the company would not let a statement like this hang in public without pushback.

That is still a guess.

The Real Question: External Safety Audit

Korbak's personal situation is not the most important thing here.

The deeper question is: does any AI company have an independent external safety audit mechanism?

The current reality:

  • Safety teams are internal departments, funded by the CEO.
  • Safety documents (Model Spec, Preparedness Framework) are written by the company itself.
  • Third-party audits (Apollo Research, METR) have no enforcement power.
  • Regulators (EU AI Act, California SB 1047) are still in the implementation phase.

In other words, when OpenAI says "we are safe," there is no external mechanism with teeth to verify it.

One possible solution: borrow the financial industry's Sarbanes-Oxley model. Public companies must have an external auditor sign off on internal controls. For frontier AI labs, that could mean:

  • Independent third-party audit required before any major model release.
  • Internal safety reports must be made public.
  • The head of safety reports directly to the board, bypassing the CEO.

These are long-term conversations. They will not happen soon.

Editorial Notes

The short-term impact of this episode: nearly zero. OpenAI's product cadence will not shift because of a single tweet. GPT-5.5, Operator, and the next agent products will ship on the existing timeline.

Medium-term impact: it further erodes the credibility weight of "former employee leaks" in AI safety discourse. Every vague leak makes the next one slightly less credible. OpenAI's PR team understands this well. Letting Korbak say this one sentence without giving more avoids legal risk while quietly depleting the "leak" weapon's ammunition.

What deserves attention is not "was Korbak pushed out or did he leave," but rather:

When an AI company's own safety head publicly says the company does not trust him, what basis does the outside world have for trusting that company's AGI safety commitments?

The difference from past leaks: Leike's accusations were specific and verifiable, which is why they got widely cited. Korbak's tweet is vague and unverifiable, so it will probably be cited once or twice and then fade. OpenAI's handling looks smarter (or worse, depending on your side).

A counter-consensus prediction: in the next six months, if OpenAI does not take new concrete governance steps (a real independent audit, full public safety reports), Korbak's tweet will become the signal moment of the "OpenAI safety culture collapse" narrative. Not because of what he said, but because of what he did not say. A classic case of a dull knife doing the damage.

What remains uncertain: no independent information cross-verifies Korbak's claim. OpenAI has not commented. If more concrete disclosures emerge (internal emails, other former employees speaking up), the picture will get clearer.

Sources

Source: https://twitter.com/tomekkorbak/status/2108266859397283953