Ex-Anthropic researcher alleges the lab is accelerating the AI self-improvement race

A former Anthropic researcher has publicly alleged that the company is accelerating the race toward AI self-improvement, according to a report by chosun.com. The claim lands on a lab that has built its public identity around safety, and it has quickly drawn attention across the industry.

The core of the report is the allegation itself: someone who worked inside the company argues that its push on self-improvement goes further than its safety positioning would suggest. At the level the headline carries, the specific projects, timelines and evidence are not laid out.

In AI, self-improvement generally means using models to improve models — letting systems take part in data selection, training recipes, evaluation pipelines and even the generation of research directions, which shortens the cycle between model generations. It is an efficient line of work and one of the hardest to bound.

What makes the claim sensitive is who is making it. Anthropic has long used safety research as a central part of its brand and its recruiting pitch, so criticism from an insider carries more weight than outside commentary and invites the reading that priorities have shifted.

Such allegations are not isolated. As spending on frontier models has climbed sharply, former researchers speaking publicly about lab incentives, release cadence and safety commitments has become a recurring genre — and the accounts usually lack details an outsider can verify.

For the industry, the real question is not one company's trade-offs but how competitive pressure systematically compresses safety processes. With compute and talent both highly concentrated, any lab that slows down cedes ground, and that structure makes restraint unusually difficult.

What to watch is whether Anthropic responds, whether the former researcher supplies more concrete material, and whether other outlets can independently verify the account. Until evidence surfaces, this remains a public allegation awaiting confirmation rather than an established fact.