Publishers Sue Brave Over the Index That Feeds AI Agents

On Monday, The New York Times sued Brave Software in the Southern District of New York, and the complaint is not really about a browser. It is about a database sitting between the open web and the AI industry. Case 1:26-cv-09549 runs 72 pages, demands a jury, and accuses Brave of building an index of more than 30 billion webpages with a crawler it never names, then selling access to that index to Anthropic, Perplexity, Mistral, xAI and Cohere. Hours later, the publishers of the New York Daily News, the Chicago Tribune and the Orlando Sentinel filed a similar suit in the same courthouse.

The language is blunt. Brave, the Times alleges, "acts both like a burglar that misappropriates the content of publishers, and an AI 'fence' that sells stolen content to third parties" according to the filed complaint. The five counts are direct, contributory and vicarious copyright infringement, a Digital Millennium Copyright Act circumvention claim, and unjust enrichment. Brave had not responded publicly at publication time as TechCrunch reported.

The Index Stopped Being a Map and Became the Product

For two decades, a search index was plumbing: you crawled, stored copies, ranked results and sent the user to the original page. Brave built its own index after 2021 to avoid depending on Google or Bing, and the complaint quotes Brave's own marketing, which claims "over 30 billion webpages, with millions of new pages indexed each day."

What changed is what Brave sells from that index. The complaint walks through three product lines in the Brave Search API: a dedicated News Search described as "a specialized index of articles sourced from trusted outlets worldwide," an AI Grounding product introduced in August 2025, and most importantly an endpoint called LLM Context. Brave's own documentation, quoted verbatim in the filing, says LLM Context "returns extracted page chunks and source metadata in a shape and length that satisfies the needs of an LLM without need for further scraping." The complaint also quotes Brave selling this to "AI agents," "RAG pipelines" and "fact checking" use cases, and its claim that the API is "trusted by the world's top AI labs," naming Cohere and Mistral and alluding to customers it cannot name.

That phrasing matters legally. A link plus a two-line snippet is a defensible search result. An endpoint whose value proposition is delivering the article body in LLM-ready chunks is a different product, and the complaint treats it as substitutive: the AI customer no longer needs to visit the publisher. The Times is not arguing that crawling is unlawful in itself, it is arguing that turning an index into a redistribution service converts a search engine into a content wholesaler with no licence.

Note where the pressure is aimed. The AI labs are named as "knowing receivers" but not sued here. The vendor in the middle is. That is the same choice News Corp made when it counterclaimed against Brave in July 2026, alleging masked crawlers and near-verbatim summaries sold to enterprise customers as Semafor reported. Brave had tried to get ahead of that fight with its own declaratory judgment suit against News Corp in March 2025 per Reuters. Two publisher groups converging on the same defendant in one year is a strategy, not a coincidence: sue the chokepoint.

The Paywall Bypass Is the Count With Teeth

Copyright cases take years and turn on fair use, which is why the DMCA claim may matter more. The Times pleads that its paywalls and anti-scraping measures are technological protection measures under 17 U.S.C. 1201, and that Brave circumvented them two ways: with automated systems that evade the measures, and by shipping a browser feature it markets as a way to bypass paywalls, described in the complaint as a default setting promoted with an explicit "Bypass Paywalls" pitch across "hundreds of top publishers."

Section 1201 is a separate offence from infringement. You do not have to prove copying or defeat fair use, you have to prove that an access control was circumvented, and the statute adds trafficking liability for tools "primarily designed or produced for the purpose of circumventing." The complaint asks for statutory damages per act of circumvention and argues each act is a separate violation, which is how these numbers get large. The Times also notes that Brave blocks publishers' ads while selling its own, so the publisher loses the impression and the reader never arrives.

Brave's likely defence is familiar from the browser wars: the user chose to view the page, and an access control that gates a business model is not the same as one that gates a work. Courts have split on how tightly circumvention must be tied to infringement, and a paywall is awkward to call a protection measure when the publisher gives articles away for search traffic. But awkward is not fatal, and a browser vendor shipping a bypass by default is a far harder defendant than a teenager with a script.

What Builders Should Take From This

If your agent retrieves from the web, you are downstream of this fight. Brave Search API keys are common in agent stacks, RAG pipelines and n8n workflows, precisely because independent indexes are cheaper and less gated than the alternatives. The lesson is not "stop using search APIs." It is that retrieval is a supply chain with a provenance problem, and that problem is now litigable.

Four practical moves:

Audit the dependency, not the vendor's brand. List every connector that pulls from a third-party index, then check what each one returns. Titles and snippets carry a different risk profile from extracted full-text chunks. If a provider advertises "LLM-ready content without further scraping," it has taken a copyright position you now rely on.

Never ship a paywall bypass. Not in a prototype, not in a scraper someone wrote on a Friday. Circumvention is a separate theory from infringement, so "we only read it, we never trained on it" is not an answer to a 1201 claim.

Identify your crawler and honour robots.txt. Both the Times complaint and the News Corp allegations lead with anonymity, the idea that an unnamed bot cannot be blocked by the publisher who owns the content. A named user agent, a published IP range and a policy you actually follow are cheap, and they remove the aggravating fact pattern that makes plaintiffs litigate.

Log provenance at chunk level and keep retrieval swappable. If you can say which URL, which retrieval time and which provider produced each piece of context, you can honour a takedown or a licence change in hours. An abstraction over providers is boring infrastructure until it is the difference between a weekend migration and an injunction.

What to Watch

Three things decide how far this goes: Brave's response and any motion to dismiss; whether the DMCA count survives early motion practice, because a ruling that a paywall is a protected access control would ripple through every scraping-adjacent product on the market; and whether Anthropic, Perplexity, Mistral, xAI and Cohere stay on the sidelines after being described as knowing receivers.

Publishers spent 2024 and 2025 suing model labs. In 2026 they are suing the data middlemen, because that is where the licence is missing and where the paper trail is cleanest. Builders should assume that "where did this text come from" will one day be asked of them, and design retrieval stacks that can answer it.