The agent ecosystem grew up fast this week. Cloudways started selling OpenClaw and Hermes like it sells web hosting, NVIDIA shipped a security stack for them, and Hermes built a one-command exit from OpenClaw. Then the uncomfortable part landed: unsealed court filings in which Microsoft’s own people called AI scraping “the largest theft of labor in human history.” Here is the week the agent economy got real, and a little scary.
Cloudways starts selling OpenClaw and Hermes like web hosting
Cloudways, the DigitalOcean-owned hosting brand, launched a Managed AI Agents product line and took both OpenClaw and Hermes to general availability. Instead of renting a VPS and wiring up gateways, ports, and security yourself, you deploy either agent from the same dashboard you already use for WordPress, with managed billing, infrastructure, and support attached.
The pitch removes the part that stops most people: not “can I run an agent,” but “who keeps it alive at 3am when the gateway drops.” When a mainstream host starts selling agents right next to web apps, running one quietly becomes a checkout decision instead of a weekend project.
Source: Cloudways launch announcement
NVIDIA ships NemoClaw so OpenClaw and Hermes stop living dangerously
NVIDIA released NemoClaw, a collection of open blueprints for running always-on agents like OpenClaw, Hermes, and LangChain Deep Agents inside a governed runtime. It wraps agents in NVIDIA OpenShell for runtime policy controls, adds model routing and local inference instead of hard-wiring a single provider, and packages a tuned Nemotron profile for teams that want domain specialization.
The framing is blunt: give the agent a sandbox and a policy layer before it starts touching your files and your shell. Runtime security has been the weakest part of the self-hosted agent movement, so it is a telling signal that the biggest chip vendor now sells that layer as the headline feature rather than an add-on.
Source: NVIDIA NemoClaw on GitHub
DeepSeek Harness shipped a no-login dashboard, and it is a CVE now
DeepSeek Harness arrived on August 13 with an impressive plugin architecture and no authentication on its web dashboard whatsoever: no password, no token, no session, no TLS. That gap is now CVE-2026-82533, an authentication bypass that grants unauthenticated access to the harness’s local agent-control API by trusting a client-supplied loopback Host header, rated critical with public proof-of-concept code already circulating.
The vendor’s own CLI is candid about the stakes, warning that binding the dashboard to the network “would expose remote code execution.” That is the vendor, in writing, calling its own control panel an RCE surface. We broke down exactly how the exposure works and how to run it safely: why the no-login dashboard is a problem.
Hermes adds “claw migrate,” and the switching war is officially on
Hermes Agent shipped a hermes claw migrate command that imports an OpenClaw setup, plus legacy Clawdbot and Moldbot configs, into Hermes. It supports a --dry-run preview, a user-data preset that skips your secrets, and side-by-side running so you can test before you commit. OpenClaw ships its own reverse migration path, so both camps are now openly poaching each other’s users.
Once switching costs collapse, agents have to compete on memory, skills, and reliability instead of lock-in, and that is a far better fight for anyone who actually runs one. We compared them head to head, including the new migration section: Hermes vs OpenClaw.
Microsoft called AI scraping “the largest theft of labor in human history”
Newly unredacted filings in The New York Times’ copyright lawsuit against OpenAI and Microsoft reveal a top Microsoft executive privately describing the companies’ AI training practices as “theft,” with the new material framing it as the largest theft of labor in human history. The same documents allege paywall bypassing, mass scraping, and copyright notices deliberately stripped from training data. Much of this comes from the Times’ own brief, so read it with that caveat.
The part that hurts the fair use defense is the internal data: Microsoft’s own numbers showed its Copilot “answer engine” cut click-through rates to the Times by as much as 93 percent versus traditional search, a decline an internal presentation called a “doom loop” that would hurt “the entire web.” When the companies doing the scraping saw the damage coming and shipped anyway, “just let models train on everything” stops looking like a neutral default.
Source: TechCrunch report
Practical: find out what AI bots can do with your site
Cloudflare also rolled out a “Disallow AI Training” setting that lets you stay visible in search while refusing to let mixed-use crawlers train on your content, with Apple, Google, and Microsoft committed to honoring the signal. If you want to know where you stand right now, the free AI Crawl Checker runs a quick check on whether your site allows AI search and citation bots (the ones that get you cited in ChatGPT, Claude, and Perplexity) separately from AI training bots, then generates a starter robots.txt you can paste straight in.
What This Means for You
If you self-host an agent, this week’s lesson is authentication. Treat every agent dashboard as a remote code execution surface, keep it on localhost or behind a real login, and patch fast, because DeepSeek Harness is unlikely to be the last tool that ships convenience before security.
If you have been putting off running an agent because of the operational work, managed hosting and hardened blueprints just removed most of that excuse. Convenience still trades away some control, so decide which one you actually value more before you click deploy.
If you publish anything, the scraping fight is now your fight too. robots.txt is a request rather than a lock, but it is a request worth making on purpose, especially now that you can separate “cite me” crawlers from “train on me” crawlers and serve each one differently.
If this week’s scraping fight has you rethinking your own site’s exposure, here is how to decide whether to block AI crawlers, and what to do instead.
Where this is heading is pretty clear. Agents are turning into infrastructure, which means the boring questions finally matter: who hosts them, what guards them, how you leave, and who gets to read your work. The companies building all of this already knew the answers were uncomfortable. The rest of us just got the filings.
About The Author
