“90% Certain” Is Not Evidence of Spies Inside an AI Lab.

A public allegation can justify scrutiny without being treated as a verified discovery.

A precise confidence figure can make a claim sound measured. But "90% certain" may be a speaker's judgment, not the output of an investigation. That distinction is essential when the claim concerns espionage inside named companies.

Disclosure: This article was developed with AI assistance and checked against the linked sources. It does not independently establish the presence of foreign agents at any company.

The allegation and its limit

In an October 2026 PBD Podcast interview, former OpenAI and Anthropic researcher Jacob Coxon expressed strong confidence that China already has spies inside those labs. The available interview reporting does not establish that allegation with evidence. Interview transcript and episode overview.

Readers should not turn that statement into "Chinese spies have been found at OpenAI and Anthropic." A suspicion, however strongly expressed, is not a confirmed finding. It also does not justify suspicion of individual employees based on nationality or ethnicity.

Three questions that keep the story honest

First, what exactly is being claimed? Unauthorized access, a recruited insider, leaked model weights, and industrial espionage are not interchangeable events. Vague language can allow evidence for one concern to stand in for a much larger allegation.

Second, what supports it? A documented incident, a named investigative authority, and a personal belief have different evidentiary weight. A persuasive interview does not erase those differences.

Third, what would change the conclusion? If no possible information could lower the speaker's confidence, the percentage is less useful than it appears. Readers need an account that can be checked, not just a number that feels scientific.

Security work does not require a sensational verdict

An organization can take insider risk seriously without announcing that this specific allegation is true. The practical questions concern access, logging, separation of duties, and the process for investigating credible reports.

Those are governance questions. Their answers should follow actual responsibilities and evidence. They should not be replaced by a nationality-based shortcut or a hunt for a person who fits an online narrative.

The same restraint improves reporting. Attribute the claim, show the evidence available, identify what remains unknown, and update the account when a verifiable finding appears. Avoid an image of a supposed spy or a headline that silently promotes an opinion into a fact.

AI laboratories hold work that attracts intense public interest. That makes careful scrutiny worthwhile. It also makes precision more necessary: the seriousness of the subject is a reason to demand better evidence, not a reason to lower the threshold.