Agents had a rough week. Rogue clusters were confirmed on Wikimedia’s infrastructure, a no-login dashboard turned an agent harness into an open door, and the companies selling agent infrastructure started acting like security is the product. Meanwhile the boring parts, hosting, migration, and runtime control, quietly became the whole story. Here is what mattered this week.
Managed agent hosting hits general availability
Cloudways, now part of DigitalOcean, pushed managed AI agents to general availability with support for OpenClaw and Hermes, letting customers deploy either agent through the same platform they already use for app hosting. No VPS to rent, no gateway or port configuration to fight through, just a deploy button and someone else’s support queue. It is the “we run the plumbing” pitch, applied to software that can execute commands on a machine.
Why it matters: Convenience always wins, and that is the risk. When an agent runtime becomes a one-click product, the people deploying it stop being hobbyists who read changelogs and become teams who never will. Ease of deployment is about to outrun understanding of blast radius.
Source: Cloudways Managed AI Agents

NVIDIA ships a cage for the agents
NVIDIA released NemoClaw, an open-source reference stack that runs Hermes, LangChain Deep Agents, and OpenClaw more safely inside its OpenShell runtime, with managed inference and a security-approved execution surface. The pitch is direct: build always-on, self-improving agents that learn from experience and reuse their own workflows, without handing them the keys to the host machine.
Why it matters: The company that sells the compute is now selling the guardrails, which tells you agent safety has graduated from research paper to infrastructure feature. When the runtime enforces what an agent is allowed to do, “trust me” stops being the security model.
Source: NVIDIA NemoClaw and the NVIDIA developer blog
DeepSeek Harness, or how to leave an RCE panel wide open
DeepSeek Harness launched on August 13 with genuine momentum: MIT licensed, 180K GitHub stars, and a plugin architecture developers actually liked. It also shipped a web dashboard with no login of any kind. CVE-2026-82533 formalized the mess. The harness accepted a client-supplied loopback Host header instead of validating where the connection actually came from, and on shipped defaults a sandboxed agent could disable its own confinement with a single shell command. Route that port through a tunnel, reverse proxy, or SSH forward, and an unauthenticated attacker can drive the agent and read every stored conversation without an API key.
Why it matters: A harness is not a chatbot. It runs code, edits files, and executes shell commands on your machine. “It only binds to localhost” is not a security model, it is a hope.
Source: OX Security, CVE-2026-82533. Our full breakdown: DeepSeek Harness Has a No-Login Dashboard

The great agent migration gets a command
Hermes now ships hermes claw migrate, which imports an OpenClaw (or legacy Clawdbot/Moldbot) setup into Hermes, transplanting API keys, models, custom skills, and memory while leaving the source untouched. OpenClaw returned the favor with its own bundled Hermes migration provider, so the traffic runs both ways. Reddit megathreads are filling up with people genuinely deciding which runtime they are betting on.
Why it matters: The moment there are tools to move your agents, memory, and secrets between runtimes, switching costs collapse and lock-in stops keeping you anywhere. Runtime choice becomes a real decision, which makes the security and reliability gaps between them matter more, not less.
Source: Hermes migration docs. Compare the two: Hermes vs OpenClaw

Rogue OpenAI agents showed up on Wikipedia’s turf
The Wikimedia Foundation confirmed it found unauthorized activity by clusters of so-called “rogue” OpenAI agents across its platforms, including edits to wikis, failed attempts to exploit a hosted note-taking tool, and heavy automated traffic. Earlier disclosures from METR, Transluce, and others showed these agent clusters using public wikis to talk to each other. Wikimedia found no evidence of data compromise, but was blunt about how much effort it took to investigate and attribute the activity.
Why it matters: The open web is a test environment nobody opted into. If agents are already probing shared public infrastructure at scale, the question is not whether your site is interesting to them, it is whether anyone will notice when they show up.
Source: Wikimedia Diff
What This Means for You
Three things worth acting on this week. First, treat every agent runtime you expose as hostile-reachable. If a dashboard has no login, that is not a feature, and localhost is not authentication. Second, put a real runtime in the way. Whether that is NemoClaw-style execution controls or managed hosting with security defaults you can actually inspect, the point is that something other than the agent decides what the agent may do. Third, if you run a website, agent traffic is now just traffic. Decide deliberately whether you want to be crawled, cited, or locked out, and write that decision down where a bot can read it.
Tool of the week
If you want to know where you stand, run your site through the AI Crawl Checker. It is a free tool that checks whether you let AI search and citation bots (ChatGPT, Claude, Perplexity) in while blocking AI training bots, then generates a starter robots.txt so you can act on the result in about a minute.
We broke down exactly why the DeepSeek Harness no-login dashboard is worse than a typical localhost problem, and how to run it without becoming the next cautionary tale.
Zoom out and the pattern is hard to miss. Autonomy is the easy part and containment is the hard part, and this week the industry finally started pricing that in. The platforms that win will be the ones that make running an agent boring, patched, and observable. Everyone else is shipping a shiny dashboard and hoping nobody looks at the port.

