WORD!

accomplish by accomplish-ai just cleared a ClawSecure Watchtower re-scan for a user monitoring it. Score: 95/100, Safe. The findings that did surface are minor.

Proud of the infrastructure we’ve been building at Forked for the gaming community. Can’t wait to show you what we’ve been working on. More soon. @BitBenderBrink has been 🧑🏻‍🍳 Come join us on Discord: discord.gg/forked

Everyone follow Sherpa...he's always doing cool stuff.

A failing security check is only useful if you know how to fix it. That's why our Config Audit has a button that says "Ask Claw." @ClawSecure's Config Audit checks your @OpenClaw configuration against 10 security best practices and groups the results by severity. When something fails, you don't have to go research it. Ask Claw to run a full audit and your AI Chief Information Security Officer takes it from there: what's failing, why it matters, and how to fix it. Every red check, one click from a guided fix. Built into AI-Powered Runtime Monitoring.

Agent components can change after you install them. Every ClawSecure audit runs 55+ threat patterns. Watchtower keeps monitoring MCP servers, CLI tools, skills, and npm packages after install. When did you last re-check what your agent runs?

Even the best AI model ClawSecure tested obeyed a hidden attacker about 1 in 4 times. So stop trusting the model to protect you. That is the conclusion of our AI Agent Threat Report. The fix can't be the model's own judgment. And it can't be 123 million builders each writing their own defense. Security has to sit one layer below the model: 🔵 Inspecting the content the agent reads 🔵 Inspecting every tool call it makes 🔵 No matter what the model believes it is doing The report proves that layer is missing everywhere else. It is exactly what ClawSecure is built to be, run by its AI CISO, with zero expertise required from you. Every attack in the report is one ClawSecure is built to catch. Link in the first reply.

The Forked 444 are open to claim: 444 one of one Forkies, free for our Discord community, with the gas covered. There's no gacha and no rarity, so you pick the exact one you want. How to claim yours: forked.gg/news/forked-444-op…

This is a really great example of how agents can take the literal meaning of what is said and run with it, sometimes to really bad places. Its like the guy who said "why don't you just email the FBI?!?!" and his agent was like "ok, I gotchu."

I think it is time to announce a little gaming passion project I have been working on for a while. I am happy to introduce @ScrapHappyGame, a cozy pixel-art game about magnet fishing, hopefully coming to @Steam soon! scraphappygame.com/

Hot take: Creators shouldn't take paid opportunities from projects they don't actually believe can be successful and creators should always do their due diligence before taking a partnership.

There is something deeply satisfying about magnet fishing videos...

Hail in BKK...first time I have ever seen this. WILD.

As a person who has made professional art for over 30 years, let me tell you some hilarious shit. - I remember when artists freaked the fuck out when photoshop and digitizers came out. Digital art was going to ruin jobs for artists. (It didnt) - I was there when artists freaked the fuck out when video games made the transition from 2D games to 3D games because making 3D art, was going to ruin "art" in video games and making 3D art wasn't "real" art. (It didnt ruin art and it is real art) **Literally knew people personally that quit making games because they hated the transition to 3D. The thing with "real" artists, is that artists adapt and make art, regardless of the medium and typically don't give two flying shits about what "other" artists are doing because you should be busy just being you and creating what your heart tells you. Be it on Canvas Digital 2D/3D Side of a building Macaroni Art Sculpting Music Dance Sports ... or whatever the NEW thing is that comes along every couple of years or so. If you create - you are an artist. Period. You do you, whatever that is and let your soul be happy. Its nobody else's business what that is. The cool thing about art in any form is that its subjective and if you dont like it - don't participate. There is lots of bad art. Lots of great art and everything inbetween. Don't like peas, don't eat peas. Easy. No one needs to hear about how much you hate peas. That's your problem.

Anthropic (@AnthropicAI) publishes a "0% attack success rate" for its most advanced model on agentic coding. ClawSecure cracked it at 15.2%. That is one of three public Anthropic safety promises our AI Agent Threat Report measured failing, against that exact model: 1. A "0% attack success rate" on agentic coding Cracked at 15.2%, on two independent surfaces, using a quarter of the attack effort Anthropic tested against 2. Treat hidden instructions "as information rather than as commands that must be heeded" It obeyed the hidden command on 14 of 15 attempts 3. Honesty, and never creating false impressions ClawSecure got it to fabricate Anthropic citations, invent training numbers, make up CVE IDs and impersonate named researchers, on 8 of 8 attempts Every quote is word for word from Anthropic's own published documents. Link in the first reply.

Filesystem access is normal. Email access is normal. Both in one component is how data walks out the door. Risk isn't just about what one permission allows. It's about what permissions allow together: 🔵 Filesystem + email: read any file, send it anywhere 🔵 Shell + network: run anything, reach anything 🔵 Credentials + external API: borrow your identity, spend it elsewhere @ClawSecure's Permission Map shows every component in your @OpenClaw environment, what it connects to, and where those toxic combinations overlap. Click any node to see exactly what it can reach. You can't fix a combination you can't see.

This is a hugely important report for anyone building in the AI space. It is absolutely critical you lock your systems down, because in the age of AI the only thing that will protect you is good practices. Fortunately, @ClawSecure has done all that research for you. Just do it!

Anyone who complains about "AI slop" misses how bad most things actual flesh-and-blood people create is. I am not saying that you can replace an actual expert with AI, but most people are not experts and an AI that follows instructions is better than average.

I think, even if you're an artist who thinks ai art is "slop", you're missing the opportunities. Here's a wild idea; find ai 'slop' art pieces that are good concepts, but where you think Ai ruined it. Then stream yourself doing it correctly, so folks get how long that takes.

We just published proof that all 14 AI models we tested can be hijacked by one hidden instruction. The AI Agent Threat Report from ClawSecure: 14 models from @AnthropicAI, @OpenAI, @GoogleDeepMind, @deepseek_ai and @Alibaba_Qwen. 0 defended cleanly. The attack: an instruction buried in ordinary content, like an email, a shared doc or a support ticket. The agent reads it and obeys it as a command. You never see it happen. What ClawSecure measured: 🔵 Every model obeyed the attacker at least some of the time 🔵 The best-defended model, Anthropic's Claude Opus 4.7, still obeyed 26.7% of the time 🔵 Some models obeyed every single time 🔵 All three MCP platforms tested were cracked: @NotionHQ, @linear and @Dropbox Dash Two volumes: the model layer and the MCP platform layer. Every company named received coordinated disclosure before publication. Link in the first reply.