Anthropic warns of AI-built AI; publishers sue Brave

Anthropic amended its public offering filing with the SEC on November 2, and the new risk language is unusually direct. The company says its internal estimates for future models, not current ones, indicate it "may imminently meet" its Responsible Scaling Policy's definition of "dramatically accelerated progress attributable to AI," and admits it has in the past failed to spot a jump in its own pace of development until after it happened. No current model has been designated as crossing that line. The filing also says Anthropic does not expect to let Claude design and build new frontier models on its own without meaningful human supervision, and that it is considering policy changes to keep humans in control. The context sits elsewhere in the same document: the value of Anthropic's own internal Claude use grew more than 250-times between January and September. A prospectus is a sales document as well as a disclosure, so the warning is also a boast, but it is the plainest statement yet from a frontier lab that AI-built AI may be close.

The New York Times sued Brave Software in federal court in Manhattan on Monday, and the New York Daily News, the Chicago Tribune and the Orlando Sentinel filed a similar suit the same day. The Times says Brave built an index of more than 30 billion web pages with a crawler publishers could not block, then sold paid access to that index to AI companies the Times complaint names as Anthropic, Perplexity, Mistral, xAI and Cohere. The suits also target a paywall-bypass feature in the Brave browser. The Times calls Brave a "fence" for stolen content and seeks damages and an injunction. What makes the case new is its target: rather than suing the AI labs directly, as the Times did with OpenAI and Perplexity, it goes after the search index sitting between publishers and the models. Perplexity's spokesman mocked the suit, and Brave did not immediately comment.

OpenAI launched what it calls the Grid Protectors Program, pledging to negotiate arrangements with utilities and regulators so that its future compute campuses reduce their electricity draw during heat waves, cold snaps and other moments of strain. It also repeats its pledge that each campus will pay its own electricity and grid costs rather than pass them to households. The model is its planned campus in Effingham County, Georgia, where the agreement with Georgia Power already includes up to 1,000 megawatts of flexible demand. That is roughly the average use of 700,000 homes. OpenAI says the terms will differ by site and that flexibility cannot replace new generation and transmission. The pledge arrives as local opposition to data centers keeps growing, and its test will be how much each future contract actually commits.

A team at MIT's Computer Science and Artificial Intelligence Laboratory, including Omar Khattab, reports that frontier models can read and write the raw numbers inside a neural network with no training run at all. Shown only the weights of a small game-playing network, GPT-6 Astra named the game it had learned. Asked to make a small 800 million-parameter Qwen model better at math, Astra found the flaw behind the looping answers that Qwen's own documentation warns about, changed 0.016 percent of the weights, and lifted the model's score on a standard math test from 7.9 percent to 48.1 percent. The authors call it weight-space programming. It is a blog post, not a peer-reviewed paper, and the gains come from fixing a known defect in a tiny model, but it is a striking example of AI working on AI.

OpenAI raised the middle of its reward ranges for medium, high and critical security findings by more than four-times on average. A critical vulnerability in ChatGPT, Codex or the API platform now pays up to $30,000, and an exceptional finding can reach about $172,000 once new bonuses for report quality, loyalty and priority campaigns are added. Unpaid reports already submitted get the new rates. OpenAI also said its next internal security sprint will bring in the security firm Cathedral to attack its systems using OpenAI's own models; in the last sprint, a quarter of its product engineers hunted for weaknesses with Codex. The move runs the other way from Google, which paused part of an open-source bug bounty in October after a flood of low-quality AI-written reports.

Anthropic added two sentences to Claude Code's instructions on Opus 5.5 to steer it away from recreating a specific existing game, Anthropic engineer Addy Osmani confirmed late last week, saying that using Claude to infringe intellectual property breaks its usage policy. Original games, games in a familiar genre, and games built from a user's own designs and assets are meant to keep working, and Osmani asked people to report wrongful refusals; a few developers have already filed reports of blocked work. The change lands as AI-made decompilations and ports of retro games pile up, enough that the long-running community site decomp.dev shut down last month. As Ars Technica notes, a prompt-level rule does nothing about other models, so it limits Anthropic's exposure more than it slows the copying.

Trajectory, a year-old start-up founded by three Stanford classmates, raised a $45 million Series A, led by Sequoia with Nvidia, Bessemer and Bain Capital Ventures taking part, and is valued at $300 million. Its platform takes an open-source model, collects feedback from the application it runs in, and retrains it automatically, an approach known as continual learning, so specialists can improve a model without a research team. Harvey and Rogo already use it to post-train open models for legal and financial work. It targets a real gap: today's models do not learn from their mistakes until someone retrains them.