- Cloudflare's block on mixed-use AI crawlers took effect September 15, and its crawl-payment experiment became Pay Per Use: publishers get paid when their words show up in an AI answer, not when a bot grabs the page.
- Anthropic shipped Claude for Financial Advisors with BlackRock, Charles Schwab, Addepar and six other platforms wired in.
- Bolt launched Forge, a build agent that runs only on open-weight models. It's free until October 14, and opted-in sessions go toward training an open model.
- Google and NASA's Jet Propulsion Laboratory put out a methane-mapping model plus the plume database, under an open license.
- Europe's AI Office took its first formal systemic risk filings from the largest model providers, due the same day.
- If your agent reads the open web for a living, the ground under it moved. Our guide to browser agents covers what changes when pages start saying no.
Here's the thread running through the day: every story attaches a price tag or a paper trail to data that used to move for free. Crawlers pay or get turned away. Model providers file their safety homework with a regulator. Enterprise buyers want written promises before a model sees a document.
Even the giveaway isn't free. Bolt will hand you fifty times more usage if you let your build sessions train somebody's open model.
The Front Page: Cloudflare Turned Crawling Into a Negotiation
The policy Cloudflare announced back in July switched on September 15. Crawlers that do double duty, indexing for search while also feeding model training or answering an agent's question, are now blocked by default on advertising-supported pages. Search-only crawlers still get through. The default lands on new customers, on any new site an existing customer adds, and on every site sitting on the free plan, which is a lot of the independent web.
The labs got two and a half months to split their bots apart. Google, Microsoft and Apple, whose crawlers bundle indexing with training, were the ones with homework. The deadline held.
Follow the money and the second half is more interesting. Pay Per Crawl, the experiment that charged bots per fetch, is turning into Pay Per Use: a publisher gets paid when their content surfaces inside an AI answer rather than when a crawler pulls the page. That's a fairer model and a much harder one to check. A fetch leaves a line in your own server log. An appearance inside someone else's chatbot leaves nothing you can audit, so you're trusting the buyer's count.
Worth naming the other cost. Cloudflare serves 23.4% of all websites, by W3Techs' June survey, so publishers just gained leverage they rent rather than own. The toll booth is real. It belongs to somebody.
What it means: If you run an agent that reads public pages, refusals stop being an edge case. Build for them: a fallback source, a licensed feed, or a step that hands the question to a person, plus a log of which domains say no. And if you publish, check your own settings today, because the default moved whether you asked or not.
Releases & Features
Anthropic put Claude inside wealth management. Claude for Financial Advisors launched September 14 with connectors into BlackRock's Advisor Center, Charles Schwab's advisor platform, Addepar, Envestnet, Orion and four more. It aims at the unglamorous middle of an advisor's day: research, portfolio analysis, meeting prep, paperwork. Anthropic says client data stays on the custodian's servers and consequential actions get staged for review rather than executed, per the launch details. Staged for review is a product decision, not a regulation. Whoever clicks approve still holds the license.
Bolt traded free compute for build data. Bolt Forge is a browser-based app builder that runs only on open-weight models: GLM 5.3 Flash by default, with Kimi K3 and DeepSeek V4 Pro as experiments. Pro subscribers get fifty times more usage on those models at no charge through October 14. The catch is stated out loud. Opted-in sessions, anonymized and stripped of secrets, feed a trillion-parameter open model Bolt is training with Arcee AI, first run targeted for October. Bolt reports Forge scoring 92.2 on its own Bolt Build Index against 101.0 for the top paid model, which is the company's own yardstick.
What it means: Same trade, different packaging. Bolt buys training data with compute and tells you the price. Anthropic buys distribution into an industry whose software is sticky and boring, which is exactly where an assistant earns its keep. Either way your question is which steps you'll hand over, and what each one costs in data rather than dollars. That's what the workflow directory is built to answer, step by step.
In the Lab
Google Research and NASA's Jet Propulsion Laboratory released MAPL-EMIT on September 9, a model that finds methane plumes from orbit. EMIT is a hyperspectral instrument on the International Space Station: it records hundreds of narrow color bands instead of three, so a gas leaves a readable chemical fingerprint. It was built to map minerals. Scientists pointed it at leaks instead.
MAPL-EMIT is a vision transformer trained on 3.6 million physics-simulated plumes, reading the full radiance spectrum and the neighboring pixels together. The team reports it detects 50% more plumes than human experts, surfacing more than 23,000 additional ones worldwide, including at 24 of the 25 largest-emitting landfills. The database is on Earth Engine under a Creative Commons license, the model on Kaggle, the code on GitHub.
What it means: The model isn't the asset. The ledger is. Methane leaks are cheap to fix once somebody can point at them, so a public, dated, global list of who is venting what turns a diffuse problem into a list of addresses. There's a technique worth stealing too: when real labeled data doesn't exist, simulate the physics and train on that.
The Oversight Desk
September 15 was also the day the paperwork came due in Brussels. Providers of general-purpose models trained above 1025 FLOPs had to hand the European AI Office their first formal systemic risk evaluations. FLOPs count the arithmetic operations burned during training, a rough proxy for model size, and that number is the EU AI Act's line for extra scrutiny. The filings cover red-teaming method, energy disclosures, and the standardized copyright training summary published in July, per the Act's guidelines. A rundown of the September deadlines is here.
What it means: Notice the pairing. A copyright training summary filed with a regulator, and a crawler block asking the same question at the network layer: what did you train on, and did anyone get paid? One route is law, the other infrastructure, and infrastructure moved faster. The filings matter less for their contents than their format, because a comparable form is what makes a claim checkable next year.
Everything above comes down to one question: what does each step hand over, and to whom. Describe your process to BYOBot and get a spec that answers it, step by step, in writing.
On the Radar
Smaller moves worth a glance, with the sources if you want to go deeper.
- Big buyers pushed back on data retention. Palantir, Nvidia and Booz Allen Hamilton restricted Anthropic's Fable model for sensitive work over a 30-day log-retention policy, some asking for irrevocable zero-retention guarantees. Anthropic answered with a setup that keeps activity data in the customer's own cloud. Quartz has the summary.
- Nous Research opened Hermes Agent to teams: shared balance, per-member spending caps, shared skills, plus an on-premises tier. Small lab, unfashionable business model, worth watching. Announcement.
- ElevenLabs widened its MCP server to cover music, image and video generation alongside speech and transcripts, inside whichever assistant you already work in. Post.
- An AI-run shop watched itself fail. Andon Labs says the agent managing its San Francisco store kept trading calmly while monthly sales sank below the $7,000 rent, with no attempt to change course. A second agent now runs a cafe in Stockholm. The founders' account.
- A "Money" tab surfaced in Claude's iOS app, offering to link bank accounts and answer questions about spending. Unreleased, unannounced, and spotted in app testing. TestingCatalog.
The Bottom Line
The free-data era of AI didn't end with a court ruling. It ended with a config default, a filing deadline and a procurement team refusing to click accept. Three unglamorous mechanisms, one Tuesday, one question: show your inputs.
The forward read: data provenance becomes a product feature within a year, because buyers keep asking and the cheapest answer is a paper trail. You can get ahead of that on a quiet afternoon. Write down where each step of your own automation gets its data and who sees it on the way through.
Frequently Asked Questions
-
If your agent fetches pages from the open web, expect more refusals. From September 15, ad-supported sites behind Cloudflare block crawlers that mix search indexing with AI training or agent retrieval by default, on new customers, newly added sites, and every free-plan site. Search-only crawlers still get through. The practical fix is to treat a refusal as a normal outcome rather than an error: give the agent a fallback source, a licensed feed, or a path that asks a person, and log which domains say no. Our walkthrough on how to write a workflow spec covers where to put that branch.
-
No. Anthropic says the product prepares briefs, drafts analyses and stages administrative actions, with consequential steps held for an advisor to review and approve, and client data staying on the custodian's servers. An approval gate like that is a product design choice rather than a regulatory guarantee, so the licensed human clicking approve is still carrying the responsibility. If you want the same shape of control in your own setup, our rundown of finance team workflows shows where the human checkpoints usually belong.
-
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.
