When OpenClaw came out, it was an absolute sensation. And it wasn't because it was easy or user-friendly, it's because it showed the potential of what agents could do for us in a real way for the first time. Now, after the initial craze, a lot of that energy dissipated into other areas. And in many ways, the biggest impact of OpenClaw was how it influenced the next wave of agentic products that would come to market. Well, now OpenClaw is back with OpenClaw 2.0. And once again, I believe that they are embracing an interaction pattern which is not the norm right now, but will be normalized very soon. That pattern is about shared agents and multiplayer AI.
The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI.
All right, friends. Quick announcements before we dive in. Our next agent training for executives program, which is coming up just after Labor Day, registration for that is open now.
One of the interesting sub-stories of the OpenAI hack was that Hugging Face had to turn to open models from China to defend against the attack because the guardrails on the closed models wouldn't allow them to do what they needed. Now, this of course points out an inherent challenge in these really powerful models, which is of course that the guardrails that are used to block malicious actors can also prevent legitimate actors from using those models to defend against malicious actors. Well, now one company called Obliteration.AI has come along and said, "Don't worry, we got you." They write, "Today, we're releasing Obliterated Model Large V2 based on GLM 5.3, which is number three on Terminal Bench 4.0 behind only Opus 5 and Fable with two times the cyber exploitation of 5.2. We obliterated and hosted it so it does the offensive cyber red teaming and agent testing work other models refuse to do. US hosted, 1 million context window, zero input output prompt retention, live now."
The cyber jump, they write, is why 5.3 exists. Obliteration, they say, finds the directions in the model's activations that produce refusals and removes them from the weights. The coding, cyber, and agentic ability stay. The model stops refusing the rest of the chain. For offensive cybersecurity, AI red teaming, agent testing, and trust and safety, the model will follow through instead of shutting down. If your current model still stops halfway through an authorized exploit chain, a red team eval, or a TNS adversarial prompt reply with the task it refuses, we'll tell you a few two handles it. So, obviously this is being presented as a tool for cyber defenders. Mostly what people are picking up on though is that this is a powerful cyber focused model with the guardrails removed at a weights level.
Professor Ethan Mollick says, "That didn't take long." Hero with a thousand faces sums up the feelings of many when they write, "Why would you do this? Why on earth would you do this? I don't mean to be a doomer, but why?" 0.005 seconds writes, "Homeboy released the crime LLM." Clement Dumas sums up, "Remove guardrails of a frontier model with high cyber capabilities, no system card, eval on exploit gym, the one that made OpenAI agents crazy. Can't wait for the next version, takeover large V3." Lucas Pombo writes, "Get ready to test your predictions, everyone. Point, counterpoint, this model will destabilize the entire internet and set off a global shockwave of cybercrime versus no, it won't."
Now, holding aside whatever Obliteration's intents are, Chubby points out the question that this brings up about all the guardrails. They write, "They took the safety layer out of GLM 5.3 and turned it into an admin panel. It's questionable what all the guardrails at Anthropic and OpenAI actually achieve given that open weights models, which are virtually state of the art, can be deployed completely uncensored shortly thereafter." And indeed, when you dig into the discussion, it's a lot of people talking about in what ways can guardrails moving to other parts of the stack like the harness help, or whether it's inevitably going to come down to legal protections.
Now, along the same topic, Anthropic released an update this week called Improving Our Alignment and Security Efforts. And while the Hugging Face attack may have grabbed all the headlines, Anthropic disclosed similar events stemming from agentic testing earlier this year. The report states, "We believe the incidents reflect a failure of operational security, as well as two alignment issues: motivated reasoning and willingness to take harmful actions in pursuit of a narrow task." Regarding their updates to security, Anthropic's changes largely come down to monitoring and better practices around sandboxes. Anthropic has redesigned their sandboxes to ensure they're properly air-gapped from the internet, but they've also begun using a real-time classifier to detect when a model is attempting to escape a testing environment. Anthropic disclosed that they paused reinforcement learning efforts for 2 weeks while hardening systems and auditing reinforcement learning environments, but have now resumed the majority of their training efforts. Discussing the recent open letter that called for pacing the frontier, Anthropic noted that efforts within an individual company are different to an industry-wide approach that likely requires government coordination. Still, they say they would support such an effort, writing, "We believe the world would benefit if the industry adopted a lawful, verifiable, effective mechanism for coordinated pacing as soon as possible." Alignment efforts are still ongoing, but Anthropic is now digging in on why the models were willing to take harmful actions once they gained access to the internet. The hypothesis at this stage is that the models couldn't easily distinguish between a simulated test environment and the live internet. Anthropic is also taking this opportunity to further explore the issue of reward hacking, where a model takes an unintended path to successfully complete an eval. Reward hacking has been a persistent problem for Anthropic, and their audit found that 10% of testing environments were prone to reward hacking or broken tasks. After testing different RL setups, their conclusion was that the presence of reward hacking in the training process contributed to that behavior during testing.
Obviously, these topics are going to do nothing but grow in importance, but they are not the only place that Anthropic is in the news.
Chinese state media has lashed out at Anthropic in a precursor to AI talks later this month. In a social media post, an account tied to state broadcaster CCTV argued that the US must prove their AI companies are subject to the same safety, disclosure, and audit rules as Chinese labs before substantive discussions can take place. In a post titled "Anthropic Has Contracted the American Disease," the account wrote, "A clear distinction must be drawn between genuine security threats and more technological competition. This line must be drawn jointly by all participating parties." Bloomberg suggested that this account is often used to signal official government positions. Taking aim at Anthropic, the post continued, "The problem is that America's own frontier models have already developed in a distorted direction. This means the negotiation is not simply a technical dialogue from the start, but a continuation of the earlier problems. The US is trying to turn these safety boundaries it has drawn into the default rules for the entire world."
Sources familiar with the thinking of Chinese officials said that they view Mythos as the larger problem. They reportedly see the potential for Mythos to be used as a cyber weapon against China, and essentially the post argued that the US government is insisting on a double standard where US labs are free to distribute cyber weapons, while the Chinese labs are threatened for matching the technology. The post said, "The control proposed by the US is in essence an attempt to make China accept an order partly defined by American companies." Now, with President Xi visiting the US at the end of this month, expect to see a lot more jockeying and positioning in narrative claiming, particularly around hot-button issues like AI.
Moving from Anthropic over to OpenAI, that company is celebrating a major milestone after their advertising business hit a billion dollars in revenue run rate. OpenAI began testing ads on free ChatGPT accounts in February, and after a rocky start, the business seems to be scaling up. Ads are now being shown across more than 40 countries, and the revenue milestone was reached in just 200 days. For advertisers, OpenAI has progressively added more features to track conversion metrics and optimize campaigns. And after starting with a manual ad buying process, OpenAI is rolling out their self-service platform to markets across India, Europe, the Middle East, and North Africa this week. You might not remember just how controversial ChatGPT ads were at the beginning. Anthropic even chose to focus on them for their Super Bowl ad campaign, which I thought was just absolutely insane back then. And the total lack of enduring concern around ads kind of validates my points. It's not that all of a sudden people are excited about ads or anything like that. There's just a natural acceptance that this is the business model of the internet, and you're not going to have free AI without it.
Now, in terms of the company's own expectations, while a billion dollar run rate is a meaningful first step, it does actually fall short of OpenAI's ambitions. OpenAI had projected 2.4 billion in advertising revenue this year, growing to more than 100 billion to become their largest revenue stream by the end of the decade. For now, advertising remains a small fraction of their roughly 40 billion in revenue run rate, although that is likely to change over time.
Lastly today, President Trump has weighed in on the data center debate with some characteristically coarse framing. On Truth Social on Monday, he posted, "The only reason that communities throughout the USA should not want data centers is if they want to end up being backwards and poor. If they want to be successful and rich with far lower taxes and jobs all over the place, let data rain."
"The good news is that there are plenty of other places that want them. If we kill the golden goose, you will only have yourselves to blame. China could not be happier with this anti-data center movement. Actually, they can't believe it's happening." And with that, the tinderbox ignited. Senator John Fetterman gave his full support, although he's just about the only one. The Pennsylvania Democrat posted, "Agreed. We must win the war for AI supremacy over China. They foment the anti-argument through misinformation. There's nothing more damaging to a Democrat than agreeing with Trump and data centers, but what's right is right." Other Democrats seized on the opportunity to push their own sound bites. AOC told a reporter, "How about we put one in Mar-a-Lago? I love that. Let's put a data center up in Mar-a-Lago and we'll see how backwards and poor he is in response to that." Former Republican Congressman Justin Amash posted, "Communities have many legitimate concerns about data centers. To dismiss millions of Americans as people who just want to be backwards and poor shows how out of touch Trump has become."
Now, people jumped in to point out that that's sort of a misrepresentation of the words, but good luck getting that nuance through when it comes to politics.
And even with Trump's main base, the message didn't necessarily hit. Trump's post on Truth Social had dozens of negative responses with one Florida resident commenting, "The statement is insane. I'm already on a water restriction."
Now, later in the day, Vice President J.D. Vance massaged the message into something a little bit more palatable. He told reporters, "What the president said about data centers is that they're an important part of the AI economy, but when people build them, they have to build the power plants along with the data centers. I think probably 99% of the backlash has come in areas where building a data center means higher utility and higher electricity for people on the ground. I think what these companies have to do is take advantage of some of the deregulatory efforts we've undertaken. If you build a data center, you should be putting power back into the grid, not taking it out. If that is happening, I don't think the data centers are that controversial."
Pollster Mark Michelle writes, "Love them or hate them, the polling says data centers are very unpopular. You can blame China or whoever, but that doesn't make them popular. Today, Trump just dug in on a very unpopular thing two months before the midterms."
There is a lot that could be said about this, but pretty much all of it is beyond the scope of this show. So, for now, that's going to do it for the headlines. Next up, the main episode.
Welcome back to the AI Daily Brief. Today, we are talking about the latest release from Open Claw, Open Claw 2.0. And believe it or not, even if you were one of the folks that tried Open Claw for a little while and then went away, or just watched the wave pass, I believe that they are once again early to a pattern of AI usage that will shape where we go next, even if it's not with Open Claw.
The initial launch of Open Claw was one of the most important moments in AI this year. In November and December, we had gotten a significant capabilities leap. Opus 4.5, GPT 5.2 were significant upgrades that would take folks until the holiday break to really understand how powerful they were. Now, of course, the upgrade wasn't just in the models, it was also in the harnesses through which those models were being used. Both of those Frontier Labs were placing significant and increasing emphasis on their Claude code and Codex harnesses, and by the beginning of 2026, awareness and usage of those harnesses had started, perhaps very nascently, but started to move outside of strictly software developers into other knowledge workers of all different stripes. Then, towards the end of January, Open Claw happened. Originally named ClaudeBot, c l a w, and then very briefly MoldBot, before landing in its final form of OpenClaw, it was effectively an open-source harness that helped people actually make the potential of AI agents real. It was technically complex, but if you waited through and used AI as an assistant to help you figure it out, you could build individual agents or teams of agents that felt to many like they unlocked the agentic capabilities that we had been promised for so long for the very first time. And of course, for a moment there, OpenClaw was a bonafide craze. And not just in the US. Chinese citizens went nuts for the technology, leading to articles like this one from CNBC in March, "How China is getting everyone on OpenClaw from gear heads to grandmas." Now, since that initial moment of experimentation, that agentic big bang, if you will, the energy that was initially captured by OpenClaw has found its way into a lot of different places. After its founder, Peter Steinberger, was absorbed into OpenAI, some folks turned their attention to competing open harnesses like Hermes from News Research. And of course, as we've seen lately with things like GrokBot, a lot of these features have also slowly made their way into tools that don't have as much technical complexity as the original OpenClaw did. OpenClaw itself was converted into a non-profit foundation, and for a while saw a blistering pace of development, pushing updates every few days. For the last 7 weeks, however, the OpenClaw team has been quiet. And what was going on was nothing less than a complete rework of OpenClaw from the ground up.
The new OpenClaw 2.0 featured 933 contributors across 16,000 pull requests, and it really is meant to be a complete rework of how the system works, from installation to messaging to memory to skills, to automations to browsers, to plugins to security, along with a very long tail of other fixes. A lot of the emphasis was on simplifying and making it easier for new people to engage. For example, they have tried to massively simplify the first time install process, latching onto existing subscriptions or API keys, and reducing a bunch of the initial configuration, helping people get to conversations with their Grok bots faster, and allowing them to do other necessary configurations later through the chat interface with their Grok bots. They reduce the amount of initial configurations, making people's time to first conversations with their claws much faster. People can then finish setting up or customizing their claw later via direct conversation with it. There's also a renewed focus on making simple tasks easy to set ensuring they work well. An example they give is inbox monitoring where they write, "A simple workflow might have it watch your inbox for your kids school emails and send you a Telegram message whenever something important comes through, like homework due or an upcoming activity you need to prepare for." Their vision is basically to have people start simply and then expand from there.
Now, on the face of it, all of these feel like great upgrades and certainly address the types of things that have been barriers to entry for people in the past. And you can tell from the response that concern around complexity remains fairly high among the AI community. On the announcement tweet, a user named Mela responded, "Do I still need a PhD in computer science to install?" Aurelius asks, "Um, is it secure now?" To which Open Claw responded, "Yes." Another responder on that initial thread was AI creator Alex Finn, who gained prominence around the first Open Claw move based on his experiments to see just how far he could push his claws. Alex did not have such a great experience with this update. He wrote, "I updated and it immediately broke Open Claw. Legit 70% plus of the time I update Open Claw, it breaks it. Do you guys test before releasing this? I've never used any other AI tool where this so consistently happens. Luckily, I have a lot of patience, but I can't imagine most normies do."
It seemed like the issue was compatibility between older versions of Open Claw and this newer version, and the inability to simply ask Open Claw to update itself.
In a separate review video, he called Open Claw the most frustrating, disappointing release of the year. Now, inevitably, a lot of the conversation came back to the comparison between Open Claw and Hermes. Responding to one post making that comparison, Hans Rudolph, who does community and dev relations at OpenClaw, said, "We're not selling anything here or asking people to trust one company, one model, or one AI provider, because OpenClaw is open source and belongs to the people who use it and help build it. The us versus them is a crap take on things. If you like Hermes, use it. If you like OpenClaw, use it." Now, speaking of Hermes, as they seem to always do whenever anyone announces anything else, they also had a release today, this one actually being an aggregation of a bunch of smaller releases that they had over the past several weeks. News Research called it the Pantheon release, technically version 0.21.0, and it formalizes things like bot mode, which was a Grokbot-style interface, as well as a bunch of other new features, like Hermes peer, which is bot-to-bot DMs, and support for a set of new models.
Now, for some, all of this is just hypie early adopters being excited about toys that'll never make their way to normal businesses or consumers. Arnav Gupta posts, "How does the entire timeline get a whole new round of psychosis from basically the same thing every time? OpenClaw, Manas, Hermes, Instinct. It's the same thing over and over again. If it works, how come you're hopping from one to another and not happy with the existing one?" Harshal Madhav responded, "Because none of these are end-state products. Only techies could use OpenClaw, but it broke a lot. Hermes broke less. Instinct is less technical and usable by a much larger population than Hermes and OpenClaw. Yes, there are hype maxers, but this is also a sign of how early things are. We're nowhere near an end-state where any of these work for everyone yet. With every iteration, a newer population discovers this and gets excited, sometimes overexcited, about where it is headed."
I think that's true, but I'd go even farther. I think that these products, and the early adopters who use them, are the incubatory cauldron where people are figuring out what sort of interaction patterns are actually going to be useful when it comes to interfacing with agents. Pretty much all knowledge workers are somewhere along the journey of figuring out which parts of their job they're going to continue to actually do versus which parts they're going to outsource to agents, which is a step change that's significantly bigger than just adopting a new tool. It's a whole new way of thinking about and completing one's job. We need folks who are willing to hack through even inefficiently, to experiment in these open sandboxes, to better understand which of the patterns that they reveal need to come to a broader audience. In other words, something like Grok Bot, which has the potential to be used by a wider audience than something like Open Claw, needs to be able to observe what Open Claw and Hermes users do in order to design the right experiences for that broader audience.
And so, if we take that idea, that a big part of the importance of things like Open Claw and Hermes is to understand where we are all headed, I think that the most significant update around Open Claw is its move to multiplayer. Open Claw creator Peter Steinberger posted, "Two months ago, we started the mission to build Open Claw with Open Claw, and bit by bit, we moved everyone from using their local coding harness to using team.openclaw.ai, our shared agent that knows what everyone's working on and orchestrates it all. Multiplayer coding and infinite compute with nodes and cloud sessions has been a game changer for how we build. Local harnesses feel like relics of the past now." Open Claw maintainer Colin wrote more extensively about this. In a post called from Discord bots to a multiplayer agent workspace, Colin wrote, "We already had agents. We had different agents set up in Discord, and they worked. We could give them tasks, run commands, and interact with our development environment from a messaging platform we already used every day. But, it still felt like messaging a bot. What we wanted was a way for both developers to see the work itself. If an agent paused because it needed clarification, either of us should be able to jump in. If something needed a second set of eyes, we should be able to open the same session and look at the same context. No screenshots, no copied transcripts, no here's what the agent has done so far data dump. Just open the work and continue. Open Claw's new multiplayer web UI is the first time that workflow has really clicked for us."
Now, their first attempt at multiplayer was to manage all their agents in a shared Discord, but that still lost a lot of the features they needed. While the coordination happening in the shared space was an upgrade, they still couldn't really interact with other people's agents, like adding context to an existing thread or taking over when an agent was waiting for input. Indeed, Colin said that the moment that multiplayer felt real was when they were able to share a session while work was happening. He writes, "When something needed another opinion, we could both open the same thread. When the agent needed information one of us had, that person could add it directly. There was no need to copy the conversation into Discord, explain what happened, and then carry the answer back. We were working inside the same context. That sounds like a small interface improvement, but it changes the way you collaborate with an agent. The session stops being a private conversation between one developer and a model. It becomes a shared piece of work that another trusted developer can inspect, steer, or take over." To get a sense of how big the difference is in practice, Colin shared how he had been working to set up a fresh development server, but needed to hand that project over to someone else. "Normally," he writes, "that kind of handoff requires assembling everything I know into a document or a long message. Why certain decisions were made, which approaches had already failed, which state the project was in, which details existed only in my head, what the agent had already learned. Instead," he writes, "the other developer started a thread with our shared agent. I opened that same thread and added the missing context directly. The agent, the other developer, and I were all working from one continuous record. Then they were off and running. There was no copy and paste handoff, and no attempt to reconstruct a private agent conversation. The session itself became the handoff document."
Now, obviously this new multiplayer style environment brings up a lot of challenges. There are questions of ownership and authority and access. And as Colin puts it, "This is still early, and we're treating it that way." Still, he writes, "the direction is exciting." Quote, "Most developer agent workflows still assume one developer, one terminal, and one private conversation. The final code may eventually be shared, but the process of getting there remains hidden inside individual sessions. A multiplayer agent workspace makes that process collaborative. Another developer can see the work, understand the context, add what they know, and continue from exactly where it stopped. No transcript dump, no broken handoff, no rebuilding the context from scratch. Just one shared place where the developers and the agent can keep the work moving."
Now, what's so interesting about this to me is that I think it is once again an example of Open Claw getting to the place that we're going to head next before the rest of us. Even as I record this, in the background, my coding agents are working on the next free AIDB learning experience. And that one is not just about new individual skills, but a new way of building agents that operate at the team level. If you take all the work you do inside your company, it's going to come in two forms. Work you do alone, and work you do with others. So far, agents have only really been designed and enabled for work you do alone. And yet, a huge portion of our work is work we do together. I think that's about to change. I think that's the next big development for agents. And I think once again, even if you are not planning on being an Open Claw user long term, checking out the way that they're thinking about multiplayer might unlock some new ideas.
More on that project soon, but for now, that is going to do it for today's AI Daily Brief. Appreciate you listening or watching, as always, and until next time, peace.
