The record labels just made personal liability the new front line in the AI copyright …

Top Story

The record labels just made personal liability the new front line in the AI copyright wars. Sony Music and Warner Chappell filed suit Friday in the Northern District of California against Anthropic — and, pointedly, against co-founders Dario Amodei and Benjamin Mann as individuals — seeking up to $150,000 per work across “tens of thousands” of songs, plus $25,000 for every instance where identifying copyright data was stripped. At the statutory maximum, the damages run to several billion dollars. But the number that matters is smaller, and it’s the one written next to Benjamin Mann’s name.

The complaint does not rest its case on the vagueness of “scraping.” It alleges a specific mechanism: that Mann used BitTorrent to download more than five million pirated books, and that Anthropic employees pulled at least two million more from Pirate Library Mirror, to build Claude’s training corpus. That is not an ambiguous fair-use dispute about model outputs — it’s an allegation of direct copyright infringement by the people who run the company, carried out through a peer-to-peer network that copyright law treats with particular hostility. The labels are also claiming that Anthropic scraped lyrics from MusixMatch and LyricFind, services that paid the publishers for the rights they then resold to Anthropic.

The suit names Amodei and Mann as defendants alongside the corporation. That is the consequential move. Until now, AI copyright litigation has mostly treated model makers as opaque corporate entities — sue the company, win damages, let the founders watch from the sidelines. By naming individuals, the publishers are asking the court to pierce that veil by the simplest route available: personal participation. If the BitTorrent allegations are credited, Mann’s individual liability is not a matter of corporate attribution at all. The named songs — Marvin Gaye and Tammi Terrell’s “Ain’t No Mountain High Enough,” Bon Jovi’s “Livin’ On a Prayer,” Earth, Wind & Fire’s “September,” Leonard Cohen’s “Hallelujah,” Taylor Swift’s “Paper Rings” — are there to make the harm concrete in a way a model’s weights never can be.

The case lands on top of a pile. Anthropic settled the book-publishing suit for $1.5 billion, and faces actions from Universal Music Group, Concord, ABKCO, BMG, and Round Hill. The music industry has effectively coordinated a siege, and the publishing settlement did not buy Anthropic the safe harbor it might have hoped for. Sony and Warner are making the calculation explicit in their filing: this is “one of the largest and most blatant ongoing thefts of intellectual property in history,” and the culprits include the founders.

What the publishers are really asking the court to decide is whether the settled framing — that training on copyrighted material is a defensible, even transformative, act — survives contact with the word “torrented.” Anthropic will argue, as it has, that even the pirated books were transformed beyond reproduction into model weights. That defense has never been tested against a founder accused of downloading five million books himself. The labels have chosen their ground carefully: not the fuzzy boundary of fair use over outputs, but the hard fact of input acquisition.

Anthropic has not yet responded to the complaint. Watch its answer for one thing in particular: whether it denies the BitTorrent allegations on the facts, or concedes the downloads and disputes only their legal significance. If Anthropic concedes the mechanism and fights on transformation, this becomes the test case for whether “scrape or pirate, it’s all fair use” survives — and its outcome will shape every model maker that has quietly accepted a torrent of training data. If it denies the facts, the discovery phase will determine whether a founder’s personal conduct, and a $2.7 trillion market’s worth of precedent, hangs on what Benjamin Mann actually did with a BitTorrent client. Sony Music and Warner Chappell are suing Anthropic

Also Today

DHS is using obscure law to snoop on journalists, non-profits, unions · Source The Guardian has the receipts on a DHS workaround that deserves to be read as a pattern, not an incident. After a judge twice rejected warrants for the YouTube data of Minneapolis journalists Georgia Fort and Don Lemon, DHS simply re-served Google with a 19 USC 1509 customs summons — no judge, just an internal sign-off, and a gag order to boot — and separately obtained six months of Fort’s phone records covering more than 10,000 calls from T-Mobile. The same summons authority was used against SEIU, the Sunrise Movement and the CWA, none of them charged with anything. Google declined to comply; T-Mobile did not. What makes this chilling is the loop it closes: a statute for customs duties is being stretched into a warrantless pipeline for first-amendment records.

Tencent Releases and Open-Sources Tencent Hy4 Preview · Source Tencent dropped Hy4 preview as an open-source model with 770B total and 49B active parameters, a 1M-token context window, and the claim of top-tier open-source performance — its internal 163-expert blind eval put it just ahead of GLM-5.3 and Kimi K3. Two details matter more than the benchmark. First, Hy4 participated in its own training and inference optimization, proposing experiments and iterating on its own output in what Tencent calls an early recursive self-improvement loop, and its self-tuned inference changes raised end-to-end throughput 31.8%. Second, the pricing — $0.834/M input, $2.501/M output — keeps the cost-efficiency play that made Hunyuan a volume favorite. The recursive-autonomy claim is the one to watch; vendors keep saying the model did the work.

Debian votes to allow "responsible use of generative AI" · Source Debian’s general-resolution vote on LLMs settled on choice 5, “Responsible Use of Generative AI”: the project neither endorses nor prohibits AI tools, but holds that every contribution must meet the same bar of quality, correctness and legal compliance regardless of how it was produced, and that the tool does not shift responsibility off the contributor. In a project that has historically been allergic to anything resembling an editorial stance, the outcome is notable for what it refuses to do. It creates no rule against AI-generated packages, which keeps Debian administratively simple, but it quietly moves enforcement pressure onto maintainers, who must now verify AI-assisted patches they may not fully understand. The compromise reads less like a policy and more like a declaration that the old vetting obligations now extend to a new class of input.

Cloudflare Workers Accept Inbound TCP, with gRPC the First Protocol on Top · Source Cloudflare finally let Workers accept inbound TCP — eight years after launch, a Worker could only be an HTTP server — via a connect(socket) handler routed through Spectrum, with gRPC as the first protocol built on top. The honest part is what Cloudflare admits in its own post: it does not use gRPC internally, preferring Cap’n Proto, and this is a private beta precisely because the platform constraint is real. Workers get unary and server-streaming gRPC through a translation layer to gRPC-web; bidirectional streaming still requires handing the socket to a Durable Object and then a Container, because fetch() cannot expose HTTP/2 stream-level control. The capability underneath is the big news — the edge can now terminate arbitrary TCP with routing logic in JavaScript — but the maturity signal is the honest one. Teams should read the private beta as a caveat, not modesty.

GrapheneOS project: pixel 11 no longer supports hardware memory tagging (MTE) · Source GrapheneOS reports its Pixel 11 port is stuck: after a week of work it holds a partial port but cannot finish it because ARM hardware memory tagging is unsupported in the software, firmware and near-certainly the silicon of the new devices. “It appears Google cut an important security feature to save money,” the project posted. MTE is the mechanism that lets the OS catch memory-safety bugs at runtime, and GrapheneOS has been the loudest champion of it as Android’s best defense against a whole class of exploitation. If MTE is genuinely absent from the Pixel 11 die, that is not a GrapheneOS problem — it is a Google decision that quietly walks back a security investment the platform had already begun shipping, and it gives every Android security argument a smaller foundation to stand on.

In Brief

• Nancy Grace Roman’s launch window opens tomorrow, sending NASA’s survey telescope toward L2 to map the sky at Hubble-class resolution. (Source)
• A Wired writer’s data-access request to McDonald’s came back as a 515-page report, while some of the other 100 companies he asked simply deleted his data instead. (Source)
• The Twelve-Factor App methodology for building software-as-a-service has been updated for 2025. (Source)
• Samsung’s processing-in-memory work keeps pressing the case that computing inside a memory chip beats moving data across a bus. (Source)
• UC Berkeley and MIT researchers released FreeToken, an open-source engine that runs frontier MoE models on consumer hardware by dynamically splitting layers between GPU and CPU. (Source)
• The OpenAI Python SDK now runs on HTTPX2 for both its synchronous and asynchronous paths. (Source)
• vLLM 0.28.0 shipped with 584 commits since the previous release. (Source)
• A security researcher describes how LLM agents, already good at navigating code, accidentally turned LLM memory into a tool for program analysis while doing vulnerability research. (Source)
• EVE Online is migrating its server code to Python 3 after more than a decade on Python 2. (Source)
• The Verge profiles musicians-turned-detectives who hunt down and publicly shame AI grifters. (Source)
• Tether brings iMessage and SMS to Linux for a switcher who found he didn’t miss much else from macOS. (Source)
• TurboKV is a fast embedded key-value store written in Rust. (Source)

One Line

It’s outrageous conduct on so many levels. It’s hard to know where to begin.

— John Roth, former DHS inspector general, in The Guardian