Skip to content

Podcast

Inside OpenClaw and Open Source Innovation

The Reasoning Show

Source ↗ ← All highlights
  • Premonition About A Social Network For AI Agents
    • Sally had a premonition about a social media network for AI agents that pushed her to study agent platforms.
    • That idea landed when Moldbook appeared and led her to dive into OpenClaw as the realization of that concept. Transcript: Sally O’Malley Yeah, yeah, sure. Okay, I’m going to talk about my premonition then. I woke up in the middle of the night about a year ago, and I had this just phrase in my head, and it was a social media network for AI agents. And at the time, I had just started learning about AI agents. And so I thought about that social media network for AI agents. I’m like, that was, that was weird, but it makes total sense because that’s exactly what we’re going to need when these AI agents are, are, are doing our business all over. So we need everything that the social media platform has, like identity management and fine grained access control and all of these things that are already in place. So I, I went and told my team and they were like, yeah, okay. And I tried to explain it. So I just filed it away. And when Moldbook came out, that was my moment to pay attention. I knew I had to dig in right there and I knew this was going to be important. (Time 0:02:38)
  • First Week Spent Containerizing And Hardening OpenClaw
    • Sally containerized OpenClaw immediately and ran it in Kubernetes and OpenShift to test deployability.
    • She added OpenShift OAuth proxy and security tweaks to validate it can run in enterprise clusters at scale. Transcript: Sally O’Malley All right. So the first thing I do when I find anything that’s interesting is I containerize it. That’s just I don’t run anything native. And so I knew that everybody was running this natively on their systems. And I want to first see if it could be containerized. There was a Docker file. So, yes, it can be. So I did that and then it runs great. And the second thing I did was try to run it in OpenShift, a Kubernetes environment. Kubernetes, no problem. OpenShift has some extra security features, some extra things you have to do to get it to work. So I did that. I put it behind an OpenShift OAuth proxy so that you have to be authenticated into the cluster in order to even access it. I did all of these little things. And I, yeah, just knew that this was such a popular project. And I also know from working on OpenShift for the past 10 years that we know better than anyone how to run applications at scale securely in a trusted way. And so that’s my mission. (Time 0:04:00)
  • OpenClaw Delivers A Customizable Personal Assistant Hub
    • OpenClaw provides a personal assistant hub you can connect to Telegram, WhatsApp, or Slack for a surprisingly magical user experience.
    • Its customization files (agent soul, identity, tasks) and scheduled jobs create a flexible, extensible assistant environment. Transcript: Sally O’Malley Exactly that. It’s it’s a way to set up a personal assistant or even an environment of personal assistance. And it’s it’s it’s easy to set up. I think it is the biggest gripe for people out there. They say it’s not, but I don’t know. And you can connect it to Telegram or WhatsApp or Slack. And so you have this AI assistant that is really engaged, you can connect with it from outside of your, your system. It’s this, you can’t even put it into words. It’s just this magical kind of feeling this aha moment when you use open claw. And that is that is why it’s exploded. (Time 0:06:09)
  • Do Lock Down OpenClaw Before You Use It
    • Do not run OpenClaw without careful configuration because connecting models and skills can be dangerous.
    • Be meticulous about access and permissions when connecting Gemini, Anthropic, or OpenAI and treat it like production software. Transcript: Sally O’Malley A bit. People are figuring it out. Think that the possibilities are uh kind of unless you can be really creative with it with what you connect it to um it uh so let me back up and mention that this project this application Is very powerful and it can be very dangerous yeah so when i when the reason I jumped into that Slack conversation was because somebody was poo-pooing it. And they were saying, oh, it’s a security nightmare. We’ll never, you know. And so I had to jump in. And I think my goal was to remind people that we have the platform that can make any application secure. And, well, as secure as it can be and so that was my i said this might be a golden opportunity for these for our platform so let’s let’s just take it and make it as locked down as as as secure As as possible yeah but it is powerful it connect it connects to the when you connect anything to the models, the Gemini, Anthropic, OpenAI, it’s extremely powerful. And so you have to configure it and you have to be really tedious and meticulous about how you give it access to anything. (Time 0:08:39)
  • Skills Marketplace Creates A High Risk Surface
    • OpenClaw’s ecosystem has thousands of skills and a marketplace, which creates a high-risk surface for downloading third‑party code.
    • Sally avoids consuming external skills and prefers to build her own until the security posture matures. Transcript: Sally O’Malley A really that’s that’s a really great analogy that i’ve thought about a lot too because in the early days of docker that’s when i started Red Hat. Yeah, everybody was downloading random crap off the internet. And you remember Dan Walsh from Red Hat saying, don’t download random crap off the internet. And so now we have this ecosystem around OpenClaw with thousands of skills, and the skills are separate. And there’s no way to, you know, it’s so new that you have to be very careful about which skills you just go and download. And I don’t download any. I make my own if I need them. Another thing is that I haven’t actually set up my open clause to be useful yet because I’m all about building, configuring, and deploying. And so I do that a hundred times a day in different ways, adding different configurations, testing different things like that’s what I’m all about. (Time 0:12:14)
  • A New Three Way Maintainer Contributor AI Workflow
    • OpenClaw development introduces a new three-way maintainer-contributor-AI paradigm where PR volume is enormous and AI assists reviews.
    • Sally described handling clusters of thousands of near-duplicate PRs and massaging contributions rather than traditional back‑and‑forth. Transcript: Sally O’Malley There was another project that we were working on called ambient and same thing is 100 ai driven and you’re just okay so uh so so yes this community this this software development process This you know the process of maintaining of of of merging prs is is different and it is the new way and we’re not going back. So currently there are 5,000 PRs open. There are clusters of PRs that are trying to do the same thing. If there’s a typo in a file, you get 20 PRs within 10 minutes fixing that typo. And so how do you as a maintainer handle that? Brian Gracely Yeah, overwhelming. Sally O’Malley Yeah, yeah, there’s, there are skills that we use and like AI skills, and they’re, they’re open source. So they’re in the in the in the repo, and you you find clusters of PRs, and you review them with AI, you review them yourself, you sometimes massage the PR rather than go back and forth with The contributor the way that, you know, traditionally we always have done. The maintainer goes to the contributor and suggests things and they get back a few days later. Like, that’s gone. (Time 0:14:20)
  • Use Rapid Community Scrutiny To Drive Security Fixes
    • Do treat the flood of contributions as an opportunity to harden security; fast community attention exposes vulnerabilities quickly.
    • Recent fixes include agent identity, SE Linux labeling, secret scanning, and an agent communication protocol to prevent spoofing. Transcript: Sally O’Malley And that’s where the focus is right now with OpenClaw, as far as I can tell, is security and stability and trying to just, you know, get it to a sane pace. Peter, it probably works like 18 hours a day and it’s not sustainable, which is why he’s bringing in maintainers to help. Brian Gracely Yeah. And Peter is Peter Steinberger, who’s the originator of the project. He’s since moved to taking a role at OpenAI, but he’s also committed to keeping the project open source and being part of the community. So, you know, we’ve talked a little bit about some of the security concerns. What are, you know, if you were to explain to somebody, and again, we don’t necessarily know where things will end up, but if you were to look at even the project that you looked at, you Know, two weeks ago, three weeks ago, when you first got started, and where you’re seeing people ask for it to be more secure. What are some of the important things that people should be kind of aware of? Or, you know, if you’re, you know, obviously a lot of projects start as sort of a, you know, a hobby or a consumer project, but eventually if it needs to get into an enterprise, a bank or An insurance company or something like, what would be some of the security things that people should be aware of that either are getting worked on or, you know, you sort of look at and You go, okay, that that’s a red flag area for now, but we’ll, we’ll see where it goes. Sally O’Malley Yeah. Yeah. It’s, it’s a million small things that are going to add up to making it robust. Just in the past few days, there’s been, um, the, I think it’s called the agent control plane. Let me make sure I have it written in my notes. Nope, it’s not. We’re going to edit that out. The ACP, it’s called the agent communication protocol, how the agents within an open claw communicate. They added an identity feature so that they can’t be spoofed. So that’s a big security feature over the last few days. Just little things like SE Linux labeling on volume mounts in POMN and Docker, improving the secret scanning. It seems like every hour there’s some security fixes going in. (Time 0:18:31)
  • Experimenting With Agent To Agent Identity And Zero Trust
    • Sally plans agent-to-agent deployments where teammates run personal OpenClaw hubs that communicate over a zero-trust bridge.
    • She’s experimenting with Keycloak, SPIFFE/Spire workload identity, and Google A2A patterns to provide cryptographic agent identity. Transcript: Sally O’Malley Actually just blown away that I’m a part of this and I’m so grateful. And I, so there’s that. And I’m really interested in running this as agent to agent. So multiple open clause, communicating with one another over our network, maybe that’s a Kubernetes cluster. Maybe it’s a semi trusted private network cluster, and you’ve got different teammates each with their personal assistant hubs which is you know an individual open claw and so me the Human I have my group of of of assistants and my teammate my agents can talk to my teammates things like that in order to make that work you you need some sort of you need identity management I’m really playing with Keycloak as a identity provider and Spiffy Spire that provides workload identity, cryptographic, zero trust identity, and the Google A to A protocol that Allows, it gives a well-known agent endpoint and you can allow your agents to communicate over this zero-trust auth bridge. There’s a project in Red Hat that’s called Cogenti that makes use of all of those things, and that’s what I’m really excited about, and I want to work to make sure that the core of OpenClaw Works well with something like that, because I think that’s going back to my premonition of a social media network for AI agents. (Time 0:22:57)
  • Expect Shadow OpenClaw Deployments In Enterprises
    • OpenClaw adoption will create many shadow deployments inside enterprises, so IT teams must proactively study and plan for identity and network impacts.
    • Brian urged IT to map how agents affect internal/external traffic, identity, and security controls. Transcript: Brian Gracely Yeah, no, it’s great to see it happening. I think if you’re curious, the good news is you can go grab this. Like you said, there’s communities building around it. There’s Reddit forums. There’s marketplaces. And if you’re an IT person, this is something to very much pay attention to because you’re going to have people that are kind of running these as shadow AI agents, just like we’ve seen Shadow stuff in the past. And if you want to start getting ahead of it, figure out, maybe figure out what people are doing with it, figure out how it’s going to impact identity, how it’s going to impact things going Inside your firewalls, outside your firewalls and so forth. (Time 0:24:28)