Cloudflare’s crawler wall went live, Salesforce shipped a CRM-specific reasoning model, and a Vite scanning campaign is quietly draining cloud keys out of dev boxes. Five stories from the last 48 hours that change something on your side.
A busy two days. Dreamforce produced the week’s most consequential model announcement, Cloudflare’s long-telegraphed deadline finally bit, and a security campaign landed that is worth an hour of your afternoon. Tap through the log below, then read the detail underneath.
[2026-09-15] cloudflare // crawler-policy
Cloudflare’s deadline landed. Mixed-use crawlers, the ones that blend search indexing with AI training and agent fetching, are now blocked by default on ad-supported pages.
Why it matters: if you run a crawler or an agent that fetches the open web, your user agent now needs to declare a single purpose or it silently stops working on a large chunk of the internet.
[2026-09-15] salesforce // dreamforce
Salesforce unveiled Koa, a CRM reasoning model post-trained from Nvidia Nemotron and running inside Salesforce infrastructure. It claims three times fewer errors than leading general models on CRM tasks, trained entirely on synthetic scenarios across 14 industries rather than customer data.
Why it matters: this is the domain-specific model thesis with a real benchmark attached. Expect your vendors to start shipping narrow models tuned to their own schema, which shifts the integration question from “which frontier model” to “whose vertical model owns this workflow.”
[2026-09-15] vite // CVE-2026-39364
Researchers disclosed a sustained mass-scanning campaign hunting exposed Vite development servers, pulling cloud credentials and infrastructure state files straight out of dev environments that were never meant to face the internet.
Why it matters: this is a supply-chain-adjacent problem you can actually fix this afternoon. Audit for Vite dev servers bound to 0.0.0.0, check egress rules on developer subnets, and rotate anything that was sitting in a reachable .env.
[2026-09-16] factory // series-d
Factory, maker of the Droid coding agents, raised 200 million dollars and tripled its valuation to 5 billion in five months. Customers named include Nvidia, Adobe, Morgan Stanley, and Palo Alto Networks.
Why it matters: the money is moving to agents that write, test, review, and deploy rather than agents that autocomplete. If your org is still evaluating AI coding tools on suggestion quality, you are benchmarking the wrong generation of product.
[2026-09-15] apple // siri-beta
Apple opened the English public beta of its rebuilt Siri, running on a custom Google Gemini model with a three-tier routing system: on-device for simple tasks, Private Cloud Compute for medium, Google Cloud for heavy reasoning. Access is waitlisted, with daily limits and a paid tier signalled.
Why it matters: the routing design is the interesting part, not the assistant. Apple is publicly validating tiered inference, cheap local model first, expensive remote model only when the task demands it. That is the same cost architecture worth copying in your own agent stack.
1. Cloudflare starts blocking mixed-use AI crawlers
As of September 15, Cloudflare blocks “mixed-use” crawlers by default on ad-supported pages. The policy, announced back in July, forces AI companies to separate the bots they use for search indexing from the ones they use for training and agent fetching. Crawlers that refuse to declare a single purpose get blocked on new Cloudflare sites and all existing free accounts. Paid customers with existing configurations keep them.
Why it matters for you: if you have built anything that fetches the open web, an internal research agent, a monitoring scraper, a competitive intelligence pipeline, its user agent string is now a policy decision rather than a formality. Check your fetch success rates this week. A quiet drop in coverage is the failure mode here, not a loud error.
TechCrunch coverage of the policy
2. Salesforce debuts Koa, a CRM-specific reasoning model
At Dreamforce on September 15, Salesforce and Nvidia announced Koa, a reasoning model post-trained from Nvidia Nemotron and built to power multi-step Agentforce workflows. Salesforce reports three times fewer errors than leading general models on CRM tasks like updating opportunities, routing cases, and scheduling follow-ups. No customer data was used in training. Instead the company generated synthetic scenarios covering tool use and reasoning across more than 14 industries. Pilot customers have it now, with general availability in winter 2026.
Why it matters for you: this is the clearest evidence yet that vertical models with narrow benchmarks can beat frontier generalists on the workflows that actually run a business. For anyone architecting agent systems, the planning question is shifting. It is less “which frontier model do we standardize on” and more “which parts of this workflow belong to a vendor model that already knows the schema.”
3. Mass-scanning campaign pulls cloud credentials out of exposed Vite servers
Researchers disclosed on September 15 that honeynet sensors recorded a sustained scanning operation targeting exposed Vite development servers. The campaign is specifically after cloud credentials and infrastructure state files sitting in dev environments, exploiting a high-severity Vite flaw tracked as CVE-2026-39364.
Why it matters for you: this is the rare security story with a same-day remediation. Vite dev servers are supposed to be local, but the default host binding and a permissive cloud security group turn “supposed to be” into “reachable from Belarus.” Three things worth doing today: inventory any Vite dev server bound to 0.0.0.0, tighten ingress on developer subnets, and rotate any cloud key that lived in a reachable .env file. Treat developer machines as production for the purposes of credential hygiene, because attackers already do.
4. Factory triples to a 5 billion dollar valuation on coding agents
Factory, the company behind the Droid coding agents, announced on September 16 that it raised 200 million dollars, tripling its valuation to 5 billion in five months. The climb went 1.5 billion in April, 4 billion in July, 5 billion now. Backers include Blackstone, Sequoia, Khosla, and Insight. Named customers include Nvidia, Adobe, Morgan Stanley, Royal Bank of Canada, Palo Alto Networks, and Ernst and Young.
Why it matters for you: the enterprise buyers on that list are not experimenting with autocomplete. They are buying agents that write, test, review, document, and deploy. If your organization is still running an AI coding tools evaluation scored on suggestion acceptance rate, the market has moved a generation past that metric. The harder questions now are review gates, audit trails, and who is accountable when an agent-authored change breaks production.
Tech Funding News on the raise
5. Apple opens the rebuilt Siri to public beta
Apple began rolling out its rebuilt Siri on September 15 as an English-language beta behind a waitlist, with daily usage limits and a paid tier signalled for later. The assistant runs on a custom Google Gemini model and uses three-tier routing: simple requests stay on device, moderate requests go to Apple’s Private Cloud Compute, and heavy reasoning routes to Google Cloud on Nvidia Blackwell hardware.
Why it matters for you: ignore the assistant, study the routing. Apple is shipping tiered inference at consumer scale, with an explicit privacy boundary at each hop. That is exactly the pattern worth stealing for internal agent platforms: classify the request cheaply, keep the easy 70 percent local, and pay for the expensive model only when the task earns it. It is also a useful reference architecture to point at when someone asks why your agent stack has three models in it instead of one.
The thread running through all five
Four of these five stories are about boundaries. Cloudflare drawing one around content. Apple drawing one between device, private cloud, and vendor cloud. Salesforce drawing one around a model that only has to be good at one domain. The Vite campaign is what happens where nobody drew one at all. If you are planning agent work this quarter, the boundary design is the architecture, and the model choice is a detail you can revisit.
If you want to go deeper on building and securing this kind of infrastructure, browse tha-shed courses.
