AI Daily Newsstand: November 7, 2026: Meta Scraping Terms, GPT-6 Cyber, Kimi K3.1 Pro

Today's moves share a shape: the brakes are being fitted from outside the buildings that make these systems. A Norwegian judge, a Brazilian court, a British competition regulator and a European compliance deadline did more to change what shipped this week than any lab's own policy did. The catch is that the one thing outside pressure can't pause is already loose. While Meta was pulling down a terms-of-service page, two companies published model weights anyone can download and keep forever.

The Front Page: Meta blinked, and it wasn't Meta's idea

Meta has indefinitely paused the commercial terms that would have let it scrape websites belonging to Meta Business account holders. Nieman Lab reported the climbdown here, and 404 Media covered how thoroughly the policy collapsed here. Meta's own wording was brief: the updates are now paused and will not take effect on November 20, 2026.

The sequence matters more than the statement. After reporters surfaced the proposal in mid-October, Meta dropped the word "train" from it, narrowed the stated use to advertising and business products, moved the start date and promised consultation. None of that stopped it. What stopped it was a Norwegian court granting a preliminary injunction after 10 media companies sued, as Press Gazette reported, a Brazilian court barring use of journalistic content without authorization, and the UK Competition and Markets Authority opening a formal investigation into whether Meta was imposing exploitative terms on publishers and advertisers. Senators Adam Schiff and Marsha Blackburn also wrote to Mark Zuckerberg asking him to scrap the terms outright.

Read "paused" carefully. A pause isn't a withdrawal, and nothing in the announcement says the underlying plan changed. Meta says it's taking its time to get this right, which is the language of a company waiting for a quieter month.

What it means: If your business has a Meta Business account, the terms you agreed to are a moving target that courts, not you, are negotiating on your behalf. Terms of service are a dependency like any API version, except they change without a release note. Worth knowing which vendors can rewrite the deal unilaterally, and what you'd do the week one tries.

Releases & Features

GPT-6 Cyber. OpenAI released its offensive-security model and, under its own Preparedness Framework, rated it Critical, one level above the High rating its predecessor carried. The company says it hits 48.5% on ExploitGym against 32.1% for the previous version, and that it surfaced 81 previously unknown high-severity flaws in Firefox. Those are its own numbers on its own benchmark, unreproduced by anyone independent. The Register covered the alignment claims and TNW the capability detail. Access is limited to vetted customers with IP allowlisting on by default.

Kimi K3.1 Pro. Moonshot AI published open weights for a 2.8-trillion-parameter mixture-of-experts model with a 1-million-token context window, activating roughly 104 billion parameters per request. Mixture-of-experts means only a slice of the model runs on any given query, which is how something this large stays affordable to serve. The weights are on Hugging Face; read the license fine print before planning a product around it.

Ripple-65B Preview. The smaller release of the day and the more useful one for most people. Ricursive Intelligence put out a 65-billion-parameter dense coding and agent model under Apache 2.0, which it says scores 80.5 on Terminal-Bench 2.1 and leads US open-weight models on agentic coding, trained on a compute budget it puts at $32 million. MarkTechPost has the breakdown. Treat the ranking as a claim and the license as the news: Apache 2.0 means you can run it, change it and ship it commercially.

What it means: Capability and permission moved in opposite directions today. The most dangerous of the three shipped behind an allowlist; the two you can download without asking went out to the entire internet. That gap is where the next two years of policy argument happens. For builders the read is narrower: an Apache-licensed coder you can host yourself is now a credible floor for a coding workflow, which changes the math on what you need to rent.

In the Lab

A new benchmark called Unsatz measures something most evaluations skip: whether a system can invent the shape of a scientific hypothesis instead of fitting numbers into a shape it was handed. An "ansatz" is the assumed form of an equation, the guess you start from before solving for coefficients. Most equation-discovery benchmarks hand that guess over free. This one doesn't. The paper is on arXiv.

Across 80 simulated systems in 20 scientific domains, nine language models solved between 21% and 60% of them, and on 44 of the 80 classical equation-discovery methods almost never succeeded at all. The sharper finding is how the models fail: they identify the right variables, get the interactions between them wrong, and when revising they favor changes that lower error on data they've already seen over changes that improve prediction on data they haven't.

What it means: That last bias is the one that makes a confident-sounding agent risky in production. A system optimizing to look right on the examples in front of it isn't a system that generalizes, and the difference only shows up once the inputs change. If you're using a model to find a pattern in your own data, the test that counts runs on next month's data, not last month's.

The Oversight Desk

Anthropic is retrofitting invisible text watermarks onto older Claude models to meet the EU AI Act's transparency rules, with a second wave reaching Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 from November 16. Models released after August 2, 2026 shipped with the mark already, and the transition period for older systems runs out in early December. The Decoder found the hole: on the second-wave models the watermark applies only while temperature, top_p and top_k sit at their defaults. Change one and the output comes out unmarked. Anthropic's own guidance tells customers who depend on the watermark for compliance to leave those settings alone.

What it means: A safeguard that switches off when you touch a routine API parameter is a default, not a guarantee. Anyone leaning on watermarking as proof of provenance should know it holds on the vanilla path and nowhere else, and that the ability to detect it sits with the vendor and regulators rather than with you.

On the Radar

Smaller moves worth a glance, with the sources if you want to go deeper.

• Meta bought Conductor. The multi-agent coding tool is joining Meta, and its cloud service shuts down November 20, after which customer cloud data gets deleted. Source.
• Apple put rogue agents in its 10-K. The annual filing now lists AI agents acting without authorization as a business risk, which is a notable thing to tell shareholders. Source.
• Nobody is letting agents book their Uber. Uber says it logged four AI-agent ride bookings in its busiest October week, plus roughly 500 agent orders on Uber Eats, against about 320 million trips and orders combined. Source.
• Long context may be cheaper to fake. A paper on compacting context reports cost savings of 3.6x to 7x with under 3.7 percentage points of accuracy loss. Source.
• A judge trimmed the Anthropic lyrics case. A federal court dismissed a copyright claim against CEO Dario Amodei personally, with leave to amend, doubting that publishers had adequately alleged Claude reproduced protected lyrics. Source.

The Bottom Line

The binding constraints on AI are arriving through courts, competition regulators and compliance deadlines, and they land on distribution and terms rather than on capability. Capability kept moving anyway, and the two models that shipped with downloadable weights are now beyond anyone's power to recall. Watch the first week of December, when the European transparency transition runs out, and watch whether Meta's paused terms return in a slower news month. Meanwhile the useful move is less about following the argument and more about picking one job you repeat every week and getting something running on it.

Frequently Asked Questions

• Does Meta pausing its terms mean my site is safe from scraping? +
  No. A pause isn't a withdrawal. The terms due on November 20 won't take effect, but Meta hasn't said the underlying plan is dead, and the court orders behind the pause are preliminary rather than final. The durable protections are the ones you control: your robots file, your contracts, and knowing which platform accounts bind you to which terms. Owning more of your own tooling helps too, which is the argument in hosted AI agents.
• If Claude text is watermarked, can I tell whether something was written by AI? +
  Not on your own. The watermark is a statistical pattern embedded at the model level, and detecting it takes tooling that sits with the vendor and regulators rather than the public. On several models getting it this month it applies only at default sampling settings, so unmarked output is easy to produce without trying. Treat it as a compliance mechanism for vendors, not a detector you can rely on. If you want AI work you can audit yourself, build it into something you can inspect step by step, which is what the workflow directory is for.
• What is the AI Daily Newsstand? +
  AI Daily Newsstand is BYOBot's daily AI news brief, published every night. It covers the day's model releases, new features and capabilities, research, and oversight news, then tells you what each move means for people building with AI, in plain English and without the hype.