OpenAI Fires Staff Over Data Shared With Outside Evaluators

OpenAI has dismissed several employees after an internal investigation concluded they mishandled sensitive information by sharing data with an outside AI evaluation group, the BBC reports. The company framed the action as a security matter. The more interesting question is what the episode reveals about how porous the boundary is between the labs building these systems and the groups meant to audit them.

External evaluation is supposed to be a check on AI developers. Independent researchers probe models for dangerous capabilities, bias, and failure modes that the makers have incentives to downplay. But that work depends on access, and access depends on data moving out of the company. When a firm can fire the people who move it, the firm controls the audit.

Key points

• OpenAI says an internal probe found employees shared sensitive data with an external AI evaluation group
• Several workers were dismissed as a result, per the BBC
• The firings sit at the tension point between corporate secrecy and independent oversight
• No regulator mandates how frontier labs handle third-party evaluations

The same week, DeepSeek shipped a native desktop client called Harness for macOS and Windows, a sign that the competitive pressure to get models in front of users is only climbing. Firms racing that hard tend to treat anything that slows them down, including outside scrutiny, as a liability to be managed rather than a safeguard to be funded.

AI lab (builds model)
       │
       │ shares data
       ↓
External evaluators     ← meant to be independent
       │
   ┌───┴────┐
 sanctioned  unsanctioned
   │            │
   ↓            ↓
 approved    firing / probe   ← lab decides which is which

The structural problem is visible in that diagram. The lab sits upstream of the evaluator and controls the pipe. If every transfer requires the company's blessing, then independent evaluation is independent only at the company's discretion. That is not a hypothetical worry when the subject matter is systems that firms themselves describe as potentially dangerous.

None of this means the fired employees were in the right. Leaking genuinely sensitive data, including other people's personal information or security-relevant internals, is a real harm regardless of the cause. But the public account is thin, and the thinness is the point. We are told there was mishandling and there were dismissals, and we have no way to judge whether the shared data was reckless exposure or exactly the kind of transparency that oversight requires.

Until there is a mandated, protected channel for external evaluation, episodes like this will keep being adjudicated inside the companies whose behavior is under review. That leaves the public trusting the labs to grade their own homework, and quietly firing anyone who hands a copy to someone else.