Fired OpenAI Safety Researchers Fight Back: 'The Reasons Look Like Pretexts'

On October 8, 2026, the OpenAI safety-researcher saga went two-sided.

Jasmine Wang, Tomek Korbak, and Mikita Balesni — the three researchers fired for allegedly mishandling sensitive information — publicly disputed the company’s misconduct claims and warned the dismissals would have a chilling effect on safety work across the industry.

OpenAI’s research leadership responded that the dismissals stand: the trio “seriously violated explicit policy” on handling sensitive information, and the breach of trust “went beyond the public letter” they had signed.

Observers aren’t so sure. Kelsey Piper wrote that the stated reasons “look like pretexts.”

The Detail That Reframes Everything

The fact pattern that makes this more than a routine HR dispute:

Tomek Korbak was OpenAI’s designated technical contact for METR and Redwood Research — during an active investigation into a security incident involving OpenAI models.

METR and Redwood Research are not random nonprofits. They are the external evaluators that:

• The White House Accord names as the model for “independent external auditors”
• OpenAI’s own Congress proposal relies on for mandatory evaluation
• The Hinton intelligence-explosion paper implicitly requires for standardized reporting

Firing your liaison to external evaluators — for communicating with external evaluators — is a message, whether or not it’s also a legitimate policy enforcement.

The Two Narratives

The researchers’ case:

• They coordinated with external safety organizations — the work the frameworks require
• The firings followed their signing of a public letter — timing that suggests retaliation
• The “mishandled sensitive information” charge is a policy veneer over the real offense: talking to outsiders

OpenAI’s case:

• The investigation found explicit policy violations on information handling
• The breach “went beyond” the letter — implying conduct, not speech, was the issue
• Companies must be able to enforce confidentiality policies or internal safety work becomes impossible to conduct

Both can be partially true: real policy violations can coexist with selective enforcement. The question isn’t whether rules exist — it’s whether they’re applied consistently, and whether these particular employees were made examples of.

The Chilling Effect Is the Point

Whether the firings were justified or pretextual, the deterrent signal is already out: every safety researcher at every lab now knows that coordinating with external evaluators can end their career.

That’s the mechanism that matters for the industry’s entire oversight architecture. The White House Accord, the Hinton paper, the labs’ own proposals — all of them assume external evaluators will have real access to internal information. If the people who facilitate that access get fired for facilitating it, the evaluators are auditing what the labs choose to show them, not what exists.

Independent oversight isn’t a document you sign. It’s a channel people maintain. October 8 made clear that channel is now contested ground.

The Pattern Since October 3

The escalation in five days:

Date	Event
Oct 3	Robinson resigns in The Atlantic: “culture is broken”
Oct 4	Three researchers fired; company cites policy violations
Oct 8	Researchers dispute claims; “pretexts”; chilling-effect warning

Meanwhile Anthropic — the lab whose IPO prospectus warns of existential risk — is moving the opposite direction: giving METR employee-level access with a no-veto publication right (more on that tomorrow). The two labs are now visibly diverging on the central question of AI governance: does “external oversight” mean access, or access minus anything that embarrasses the lab?

What This Means for Business Leaders

1. Vendor Safety Culture Has a Public Scoreboard — Track It

Who’s leaving, who’s being fired, whether evaluators actually get access. These are now leading indicators of internal safety posture, and they’re all publicly visible. Build them into your vendor risk reviews — a lab hemorrhaging safety staff while firing evaluator liaisons is telling you something about what the evaluators would find.

2. “External Audit” Claims Need a Verification Question

Every lab now claims independent evaluation. The follow-up question that matters: do evaluators have unrestricted access, or curated access? The Wang-Korbak-Balesni fight is the exact scenario that distinguishes the two. Ask vendors directly — and note how they answer.

3. Whistleblower-Adjacent Risk Is Now a Deployment Risk

If your AI deployment depends on a vendor’s safety claims, and that vendor’s safety staff are being fired for coordinating with evaluators, your assurance chain has a gap. Diversify verification: independent evals where available, contract clauses covering evaluation access, and incident-reporting obligations that don’t rely on the vendor’s internal culture surviving contact with scrutiny.

4. Watch Which Lab’s Model the Market Adopts

Anthropic’s METR access vs. OpenAI’s researcher firings is a live A/B test in safety-governance postures. Enterprise adoption will follow whichever approach produces fewer incidents and better disclosures. The Executive AI Workshop practice tracks these divergences for exactly this purpose — the governance model your vendor uses predicts the incidents you’ll inherit.

The Bottom Line

The OpenAI researcher fight has moved past “were they fired fairly” into “can external AI oversight exist at all.” If the people connecting labs to evaluators can be fired for the connection — whatever the formal reason — the oversight mechanism every framework depends on is only as strong as each lab’s tolerance for embarrassment.

For businesses, the diligence translation is concrete: when a vendor says “independently evaluated,” ask whether the evaluators saw everything or just what was shown. The answer now determines whether your AI assurance is real or performative — and this week’s dispute is the case study that proves why the question matters.