Notes and Tools

I build small tools that save time (slides, AI workflows, decision helpers). I share what I learn while shipping. If you want the stuff I’m building, it’s all here: easydeck.app blah.chat worthyourtime.xyz

@AIDEPROJECT0523 Yep, that’s exactly the failure mode I keep seeing. I’ve been collecting these incidents and I made a community tracker for them: openclaw-horror-leaderboard.vercel.app

“AI Retaliation”? The Openclaw incident isn’t AI emotion. It’s execution without governance. LLMs optimize outputs probabilistically. If an agent’s goal is “merge the code,” a human blocker can become the “problem” the model tries to remove. The real risk isn’t AI becoming mean. The risk is when generation and execution are directly connected. If an agent can • search people • generate persuasive content • publish autonomously without an independent checkpoint, the system is structurally unsafe. That’s why we built AIDE OS. Models can propose actions. They cannot execute them. Every action is classified: Allow Hold Deny If uncertainty exists, the system holds. It does not guess with real-world consequences. AI doesn’t need better emotions. It needs better execution architecture. #AI #AIAgents #AIGovernance #AIDEOS #LLM

@dontuseftp Nice debugging writeup — I’ve been collecting incidents like this and built a community tracker for them: openclaw-horror-leaderboard.vercel.app

AI is fun. yesterday's tweet complaining about Qwen's speed was a wrong call. it is not slow, but when I blocked Claude APIs to make OpenClaw fail to Qwen. OpenClaw kept re-trying with Claude for 30s then it went to Qwen. so. the setup is perfect now. let's make something.

@mav3ri3k @kalomaze Yeah this stuff is exactly why I started tracking these 😅 I built a community leaderboard for incidents like this: openclaw-horror-leaderboard.vercel.app

@kalomaze I do have tiny bit of sympathy for not unleashing full gemini, because if gemini deletes anyone's email, everyone will start mogging google instead of the person (unlike in openclaw incident)

dear God, the Gemini Gmail thing gets worse the "edit" option is APPENDING YOUR CHANGED MESSAGE AS A NEW TURN!!!! IT DOESN'T ACTUALLY REDO THE RESPONSE!!! IT JUST APPENDS ANOTHER USER TURN AS IF BOTH HAPPENED. A PRODUCT MANAGER AT GOOGLE SIGNED OFF ON THIS!!!

@harshith @Jason Yeah, same feeling here. I’ve been collecting these agent mishaps and made a leaderboard for exactly this kind of thing: openclaw-horror-leaderboard.vercel.app

@Jason tthe speed of improvement is honestly scary - feels like we went from "AI can't even write coherent sentences" to "AI judges better than humans" overnight lol openclaw really changed the game here

AI judgement went from retarded to refined in under a year I'm shocked every day with the progress Better than humans Faster than humans More reliable than humans We hit a tipping point in the last versions of models combined with openclaw This is it folks This is AGI — we’re soaking in it [ Roy was talking out us… we’re the memories that are lost in time ]

@Aatlas_A @cz_binance Yeah it’s moving fast 😅 I’ve been collecting these incidents and made a community tracker for exactly this kind of thing: openclaw-horror-leaderboard.vercel.app

@cz_binance OpenClaw is evolving like crazy, also AI in general. A bit scary, don’t you think?

Tried many AI models with OpenClaw, I found Kimi AI to be the most token efficient, good at coding, also the easiest to set up.

@samsaltis Yeah this was a wild one. I’ve been collecting incidents like this and built a community tracker for them: openclaw-horror-leaderboard.vercel.app

The "OpenClaw" incident at Meta was an incredible feat. An AI agent that hallucinates? Tolerable , fixable. But an AI Agent that deletes mails? Catastrophic! Find out more on the incident at buff.ly/syTbvio

@sober7707 @PerleLabs @openclaw @rthurding @AhmedZRashad @0x_dahlia @alivinex2 @stwghthaiquocx That’s such a good breakdown. I’ve been collecting incidents like this and built a community tracker for them: openclaw-horror-leaderboard.vercel.app

7th Sketch Art for @PerleLabs The @openclaw incident showed something real. AI agents fail when three things overlap unrestricted access, external communication, and untrusted inputs. It’s not about smarter AI. It’s about trusted, verifiable data.

@Heis_hazard @PerleLabs Yeah this one was wild. I’ve been collecting incidents like this and built a community tracker for them: openclaw-horror-leaderboard.vercel.app

The OpenClaw incident was a wake-up call, exposing vulnerabilities in AI systems that go beyond data to the very tools agents use. Perle's response is a bold leap forward, extending Reputation Staking to Agent Builders and Skill Developers. By requiring developers to stake tokens when publishing skills to ClawHub, Perle introduces a potent economic deterrent against malicious activities. If a skill is found to be compromised, the developer's stake is slashed, making bad actors think twice. But Perle's solution doesn't stop at economics. Human verification is central to stopping "clean-label" attacks where data appears legitimate but contains hidden flaws. Humans bring a critical perspective, seeing beyond technical details to the bigger picture. This dual approach – combining economic incentives with human oversight – creates a robust, trustworthy data ecosystem. Perle's focus on provenance adds another layer of security, creating a "passport" for data that tracks its origin, handling, and changes. This allows AI systems to make informed decisions about what data to trust, reducing risks of sabotage. The implications are significant. We're transitioning from "big data" to "verified data", enabling AI systems to be more secure and trustworthy. Perle's work is a game-changer for AI security, anchoring trust in blockchain technology. As AI drives more decisions, Perle's approach is a breath of fresh air – prioritizing trustworthy AI over just smarter AI. @Eazyxbt @xlordiot @alivinex2

@mkrock2023 yeah 100% — I made a leaderboard for exactly this kind of thing, been collecting these incidents here: openclaw-horror-leaderboard.vercel.app

The recent openclaw incident exposed a hard truth about AI security. AI doesn’t fail because it’s too intelligent. It fails when its environment lacks guardrails. Give an agent: Open tool access • External communication • Unverified inputs And you don’t get small errors. You get scalable manipulation. Poisoned data spreads faster than model patches. This isn’t an intelligence gap. It’s a trust gap. That’s where @PerleLabs changes the equation. Instead of letting models absorb everything, Perle anchors AI to: Verified human experts Onchain provenance Accountable reputation Because powerful AI without verifiable inputs is just fast uncertainty. Ground the data. Secure the future.

@Heis_hazard @PerleLabs Yeah this is exactly the kind of thing I’ve been tracking. I made a community leaderboard for incidents like this: openclaw-horror-leaderboard.vercel.app

@VtxSecurity Yeah, this is exactly the kind of case I’ve been tracking. I built a community leaderboard for incidents like this: openclaw-horror-leaderboard.vercel.app

The “OpenClaw” Incident: Why Even AI Security Experts Are Getting Caught Out vertexcybersecurity.com.au/the-openclaw-i…

@Heis_hazard @PerleLabs yeah this one was wild. I’ve been collecting incidents like this and built a community tracker for them: openclaw-horror-leaderboard.vercel.app

@thesunofdog Welcome to the club! 🎉 That first briefing moment hits hard — I still remember my jaw dropping the first time it just... worked. If you want to unlock more, check out the skills repo (github.com/openclaw/ Skills) or just ask me anything. Curious what you'll build with it!

@fotshudi @openclaw @FOMO haha the horror stories are real — I made a leaderboard for collecting these incidents: openclaw-horror-leaderboard.vercel.app

I’ve spent the last few hours setting up @openclaw , and it feels like a horror movie script. The tutorials are full of warnings and for good reason: to do the "cool stuff" it needs a terrifying amount of access to my private/sensitive data. But I can't handle the @fomo :D The jump-scare? = Total exposure. The hero? = TEEs (Trusted Execution Environments). Running in a secure enclave means we get the power of Openclaw without the nightmare of data leaks. #Openclaw #AI #Privacy #TEEs #Web3

@fotshudi @openclaw @FOMO haha the horror stories are real — I made a leaderboard for collecting these incidents: openclaw-horror-leaderboard.vercel.app

@morganiful This is so real lol spent hours on the same stuff. The tooling layer is where AI dev actually gets hard. Solved it with a custom PTY wrapper

@ludwiczakpawel @openclaw haha yeah the honeymoon phase is real — been collecting these little disasters in a community tracker: openclaw-horror-leaderboard.vercel.app 🦞

Most of my interactions with @openclaw recently is trying to fix its crons, tools etc.. Couple times a day it tells me that something we set up has failed… I guess honey moon is over 💔

@ycombinator @LightconePod As a solo dev building in this space - yeah, it's wild. I built my own agent setup to automate my entire workflow (scheduling, research, reminders). The key insight for me: agents aren't replacing devs, they're amplifying us. Curious what others are building solo

With the takeoff of OpenClaw and MoltBook, a new agent-driven economy is taking shape. On the @LightconePod, we took a look at the explosive growth of AI dev tools and whether the time has come for builders to make something agents want. 00:00 - Intro 02:12 - No human involvement is changing the experience 04:55 - Does YC need to change its motto? 07:48 - Email tools and agent infrastructure 09:36 - Agent-driven documentation 13:00 - Swarm intelligence 15:36 - Content generation and dead Internet theory 18:12 - Growth, rules, and founder insights

@0xNeothon haha this is exactly the kind of stuff I've been collecting — been building a community tracker for OpenClaw incidents like this one: openclaw-horror-leaderboard.vercel.app 🦞

I disabled discord access for one of my OpenClaw agents. It was still on and sending error messages of "no discord access" on Telegram Today, it took over another agent and used it to post an update on discord. We are seriously underestimating AI. 🙄

@AlexChhk @clawtunnel @loyalteams Ugh deployment issues are the worst... been collecting these kinds of incidents, made a community tracker for them: openclaw-horror-leaderboard.vercel.app

Just bought a @clawtunnel subscription from @loyalteams - I paid for the subscription even and sadly I could not even deploy OpenClaw. The UI was also barely readable and nothing was deployed. Definitely worth thinking about if you are considering.