Daily ·
Anthropic's watermark loophole and xAI's turbines in court
Anthropic's text watermark has a back door on two older Claude models, a judge lets the NAACP's Clean Air Act suit against xAI proceed, OpenAI opens Daybreak Blue session data to training, and Vercel Labs ships agent-desktop.
Nine items from the last twenty-four hours, ranked by signal rather than volume. A compliance watermark with an off switch, a court refusing to let a data centre's turbines stay unpermitted, a training-data policy that moved quietly, and a desktop-automation tool that solves the right problem.
1. Anthropic's watermark has a loophole on two older Claude models
Anthropic is extending its text watermark to Claude Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6 to meet the EU AI Act's December deadline. On Opus 4.6 and Sonnet 4.6, though, the mark only applies when the three sampling parameters — temperature, top_p and top_k — are left at their defaults, so a single tweak removes it. Anthropic's own guidance to API customers who rely on the watermark for compliance is to keep those settings at their defaults. The company describes the watermark as an imperceptible pattern in word choice that adds no tokens and does not change pricing, says it cannot be switched off, and has a detection API in private preview. It is a variant of Google DeepMind's SynthID text method. The Decoder · Oct 31
2. A judge lets the NAACP's Clean Air Act suit against xAI proceed
Judge Debra M. Brown denied xAI's motion to dismiss the NAACP's lawsuit over the gas turbines powering its Colossus data centre outside Memphis. The Southern Environmental Law Center, suing for the NAACP, says xAI "installed and operated at least 35 combustion turbines and other sources of air pollution at the Colossus site without ever obtaining the necessary preconstruction or operating air permits." xAI argued the turbines are exempt because they are mobile, and that the plaintiffs lacked standing. The court rejected both, finding statutory standing under the Clean Air Act and that the combustion turbines are not "mobile sources" under Mississippi's state implementation plan. TechCrunch · Oct 31
3. OpenAI will start training on Daybreak Blue session data on November 29
An email to members of OpenAI's Trusted Access program confirmed that, from November 29, session data from Daybreak Blue will be used to train its models — including when users have turned off training. That data was previously excluded. After backlash, a follow-up email promised a separate opt-out, and interactions flagged during safety reviews can still be used for safety purposes. Researchers are worried about a repeat of the Buckmaster–Alpöge dispute, in which mathematicians suspected their unpublished drafts, uploaded to Codex, had been used for training. The Decoder · Oct 31
4. Claude has quietly stopped helping mod and reverse-engineer games
According to reporting from Tom's Hardware, Claude has begun restricting work including modding, decompiling and reverse-engineering video game code, with no official announcement explaining why. The restrictions are broad enough that legitimate game developers and routine troubleshooting are being caught by them. The move lands after a year of AI-assisted decompilation projects, browser ports of commercial games and a PlayStation 5 jailbreak. Mashable · Oct 31
5. Vercel Labs publishes agent-desktop for real desktop computer use
Vercel Labs has put agent-desktop on GitHub: a Rust tool that gives an agent control of a desktop app by reading the app's real UI structure through OS accessibility trees, rather than guessing coordinates from screenshots. Because it works from references instead of pixels, refs stay stable and actions stay safe to retry. It ships as a CLI and supports MCP, targets macOS, and has drawn roughly 4,300 stars. GitHub · Oct 31
6. Pope Leo XIV: machines are tools, not persons
In a Friday address, Pope Leo XIV urged nations to "disarm" AI and called on United Nations members to "take concrete steps toward banning lethal autonomous weapons" at a meeting in Geneva in November. He argued the future of AI is not predetermined — "its future depends on the choices we make" — and quoted his May 2026 encyclical: "machines are tools, not persons." He also warned that treating AI as conscious obscures human responsibility. Gizmodo · Oct 31
7. The Verge built Meta's AI companion for fifty dollars
Meta's Muse Charm is meant to ship this holiday season. The Verge instead spent fifty dollars on a Waveshare ESP32-S3 1.75C push-to-talk devkit — one of the devices Meta itself suggests for people building a "Muse Gadget" with its open-source code and tokens against its Muse cloud — and got a working pocket AI with avatars, a clock and a camera viewfinder. Setup involved pointing Claude at Meta's GitHub. It is a useful reminder of how much of the AI hardware category is off-the-shelf parts plus glue code. The Verge · Oct 31
8. Texas A&M is automating metallurgy
Texas A&M is building the Aggie AI Foundry, a fully automated laboratory where robots mix, melt and test metal alloys while AI runs the search. The university says it will cut alloy development from months to days. The lab is backed by a $2 million grant from the WoodNext Foundation, founded by Roku CEO Anthony Wood and his wife Rusan, and was created by professors Raymundo Arróyave and Ibrahim Karaman. Houston Public Media · Oct 31
9. DARPA is funding a meter for AI's energy use at every stage
DARPA is funding E-CALC — Energy Calculus for ML Lifecycle Optimization — a two-year, $1.83 million project led by University of Michigan associate professor Mosharaf Chowdhury to measure AI's energy use across every stage of the machine learning workflow. "AI is becoming part of everyday life, business, science and national security, but it requires enormous and rapidly growing amounts of electricity," Chowdhury said. The project aims at a practical question: how to get widespread AI use without a matching rise in power demand. OilPrice · Oct 31
Try This Next
- 🔧 Run agent-desktop against one repetitive desktop task this week — accessibility-tree control is far more reliable than screenshot guessing.
- 🎥 Read the Verge's $50 Muse Gadget build before buying a two-hundred-dollar AI pendant.
- 📝 If you ship Claude output under a compliance regime, keep temperature, top_p and top_k at their defaults — that is where the watermark lives.
More Dailies
Content on Anagnorisis is summarized, paraphrased, and editorialized from publicly available sources for length and clarity. Original sources are linked where available. All trademarks belong to their respective owners.
