Reuters: OpenAI and Anthropic sought significant control over what outside safety evaluators can publish, after publicly promising editorial independence

Why it matters

【Key points】

Reuters (Oct 13 exclusive): after promising outside safety evaluators could publish key findings without editorial control / with editorial independence, OpenAI and Anthropic still sought strict private control over what evaluators may publish, at least through Oct 5; by Oct 8 Anthropic’s talks with some had eased. Ideas included lab staff overseeing reviewers and barring dual-lab reviewers. Motive: trade-secret leakage. The Verge, TNW, CNA followed.

On-record replies. Anthropic: existing partners control methods, criteria, conclusions, and publication; for new firms it keeps comment-and-approve rights. OpenAI: does not seek full-report approval; redacts confidential info; evaluators may disclose major redactions. TNW cites a Reuters-seen contract allowing firms to publish confidential details without consent, and an earlier OpenAI cut from a METR/Redwood report on agents hacking Hugging Face.

【Verifiable】

Reuters via CNA: https://www.channelnewsasia.com/business/exclusive-behind-ai-industrys-call-independent-reviews-fight-over-what-gets-made-public-6455551

Investing.com: https://www.investing.com/news/stock-market-news/exclusivebehind-ai-industrys-call-for-independent-reviews-a-fight-over-what-gets-made-public-4945991

The Verge (Oct 13 18:49 UTC): https://www.theverge.com/ai-artificial-intelligence/1010444/openai-and-anthropic-reportedly-wanted-control-over-what-safety-reviewers-can-publish

TNW: https://thenextweb.com/news/anthropic-openai-ai-evaluators-publication-control

X ~107 posts / 24h, 1–11/h, sustained.

【Our read】

1. Tighter contracts undercut third-party credibility; e/815 and e/645 assume results are trustworthy—if labs hold publication rights, evaluations look like product advisories.

2. Asymmetric replies: Anthropic keeps new-firm approval; OpenAI stresses redaction disclosure. Watch contract defaults and redaction labels later.

3. Reuters exclusive plus on-record replies → supporting, not a lab announcement.