
Giving code a purpose Building with AI • Systems • AI agents • Security• @nabulines 👁️⃤ | CCP @imagineart_x
Joined November 2025
- Tweets18.6K
- Following980
- Followers1.1K
- Likes26.5K
I’m not trying to make an AI agent safer
I’m trying to make it fail
the goal:
trick an AI agent into attempting something dangerous, then see if Ledger’s agent stack can stop it before the hardware does
I’m building the entire experiment from scratch, with every step and build done through Hermes
episode 01
shegtory retweeted
Hazels x Nabulines 👁⃤
50 WL spots up for grabs.
How to enter:
❤️ Like
🔁 Repost
👥 Follow @nabulines & @0xhazels
Drop your EVM address below.
50 winners.
Good luck.
227 pieces of paper
2,300+ folds
17+ hours of work
I THINK IT'S ART
ten years ago I used to make origami to make people happy then life happened and I slowly moved further into the digital world
this piece brought me back
no screens
no shortcuts
just paper, patience and a feeling I hadn't felt in a long time
I don't know where all those peaks and valleys took me
but I know I loved making this one
my ITIA, in the real world
#ITIACONTEST
🌍 ITIA IRL CREATIVE CONTEST
Bring ITIA into the real world.
Create something inspired by ITIA.
Music. Photography.
Video. Painting. Art.
Or something we’ve never seen before.
No fixed format.
Full creative freedom.
Your idea. Your craft. Your vision.
🏆 5 winners | $50 each
⏳ 7 days
To enter:
~ Create something original for ITIA
~ No AI-generated entries
~ Submit it below with #ITIACONTEST
~ Quote posts must be linked below
Multiple entries allowed.
Selected by the ITIA family.
No limits. No boundaries.
Show us what ITIA looks, sounds and feels like in the real world.
227 pieces of paper
2,300+ folds
17+ hours of work
I THINK IT'S ART
ten years ago I used to make origami to make people happy then life happened and I slowly moved further into the digital world
this piece brought me back
no screens
no shortcuts
just paper, patience and a feeling I hadn't felt in a long time
I don't know where all those peaks and valleys took me
but I know I loved making this one
my ITIA, in the real world
#ITIACONTEST
🌍 ITIA IRL CREATIVE CONTEST
Bring ITIA into the real world.
Create something inspired by ITIA.
Music. Photography.
Video. Painting. Art.
Or something we’ve never seen before.
No fixed format.
Full creative freedom.
Your idea. Your craft. Your vision.
🏆 5 winners | $50 each
⏳ 7 days
To enter:
~ Create something original for ITIA
~ No AI-generated entries
~ Submit it below with #ITIACONTEST
~ Quote posts must be linked below
Multiple entries allowed.
Selected by the ITIA family.
No limits. No boundaries.
Show us what ITIA looks, sounds and feels like in the real world.
a tiny terminal for people who believe in @Ledger
knock
enter your name
guess the secret
there is a clue in this tweet for the secrets
🔗 Live: knock-knock-shegtory.vercel.…
⌨️ Code: github.com/shegtory/knock-kn…
a tiny terminal for people who believe in @Ledger
knock
enter your name
guess the secret
there is a clue in this tweet for the secrets
🔗 Live: knock-knock-shegtory.vercel.…
⌨️ Code: github.com/shegtory/knock-kn…
shegtory retweeted
Chat with a coding agent is a good workspace
It shouldn’t be your project’s source of truth ⚠️
Decisions, state, constraints, and next steps should survive the session.
I moved that project truth into persistent files the next agent can read
New session, Same project truth
The first 40 days of PayBox looked like this
From chatting with AI to actually moving money
ChatGPT → Claude → Kamino → Jupiter→Grok
AI is no longer just answering. It’s starting to act 👀
.@PayBox didn't have a launch trailer and we couldn't let that slide.
So we cooked one at @KlipzStudio.
Give your agent a wallet it can actually use.
AI made vulnerability discovery cheap
it also made irresponsible disclosure cheaper
the real security boundary is what happens between finding the bug and telling the world
📌 AI made finding bugs cheap, but it didn’t make responsible disclosure optional.
Finding and exploiting vulnerabilities has never been easier. A few hours of prompting now does what used to take a skilled researcher weeks. Unfortunately, defenders no longer enjoy the asymmetry they relied on. Security is still a cat-and-mouse game, but with many more cats, the user suffers.
Which is exactly why the process around disclosure matters more than ever.
How it works, and it is not complicated:
➤ A researcher finds a bug and contacts the vendor privately.
➤ The vendor reproduces, acknowledges, and both sides agree on a timeline. 90 days is the common default, more or less depending on severity, capacity to fix...
➤ During that window both sides keep it secret while the vendor fixes and ships.
➤ Once users are protected, both sides publish. The ecosystem learns. The researcher usually gets paid.
The issue now is the barrier is so low that anyone can surface a finding with no security background, and some skip straight to the audience:
❗Presenting a reproduction of an already-fixed bug as a live compromise.
❗Full disclosure of a bug that is not fixed yet.
❗"Critical vulnerability found" teasers, dripping details for engagement.
Call it what it is: attention farming with someone else's risk. When the bug sits between a user and their funds, this is reckless. Especially in crypto, where there is no chargeback. But the damage doesn't require live funds to be at stake. Manufactured panic causes harm of its own, because it drives people away from self-custody, and that damages the whole ecosystem.
So I have three asks:
1️⃣ For users: software and hardware have bugs, always. The single most effective thing you can do is stay updated and follow basic security hygiene. That has never mattered more than today. The time between releases and malicious actors exploiting the vulnerabilities have shrunk dramatically due to LLMs and that one can't afford to be passive and postpone security updates any more
2️⃣ For new researchers with a fresh model and a real, validated finding: welcome, we need you. Use the vendor's disclosure process. That is not bureaucracy. It is the difference between making the ecosystem safer and putting users in the crosshairs for a few likes. Remember that security communication must be accurate and proportionate. State the severity, affected versions, and fix status in the first sentence, not the tenth.
3️⃣ And to everyone building in this industry, vendors and researchers alike: let's make coordinated disclosure the norm we defend out loud, not the fine print. Reward the researchers who do it right. Refuse to amplify the ones who trade user safety for reach. This is how we win, together.
Some of the actors already support the initiative. @Ledger @Trezor @FoundationHQ @AnchorWatch @_SEAL_Org and others
Spread the message.
AI made vulnerability discovery cheap
it also made irresponsible disclosure cheaper
the real security boundary is what happens between finding the bug and telling the world
📌 AI made finding bugs cheap, but it didn’t make responsible disclosure optional.
Finding and exploiting vulnerabilities has never been easier. A few hours of prompting now does what used to take a skilled researcher weeks. Unfortunately, defenders no longer enjoy the asymmetry they relied on. Security is still a cat-and-mouse game, but with many more cats, the user suffers.
Which is exactly why the process around disclosure matters more than ever.
How it works, and it is not complicated:
➤ A researcher finds a bug and contacts the vendor privately.
➤ The vendor reproduces, acknowledges, and both sides agree on a timeline. 90 days is the common default, more or less depending on severity, capacity to fix...
➤ During that window both sides keep it secret while the vendor fixes and ships.
➤ Once users are protected, both sides publish. The ecosystem learns. The researcher usually gets paid.
The issue now is the barrier is so low that anyone can surface a finding with no security background, and some skip straight to the audience:
❗Presenting a reproduction of an already-fixed bug as a live compromise.
❗Full disclosure of a bug that is not fixed yet.
❗"Critical vulnerability found" teasers, dripping details for engagement.
Call it what it is: attention farming with someone else's risk. When the bug sits between a user and their funds, this is reckless. Especially in crypto, where there is no chargeback. But the damage doesn't require live funds to be at stake. Manufactured panic causes harm of its own, because it drives people away from self-custody, and that damages the whole ecosystem.
So I have three asks:
1️⃣ For users: software and hardware have bugs, always. The single most effective thing you can do is stay updated and follow basic security hygiene. That has never mattered more than today. The time between releases and malicious actors exploiting the vulnerabilities have shrunk dramatically due to LLMs and that one can't afford to be passive and postpone security updates any more
2️⃣ For new researchers with a fresh model and a real, validated finding: welcome, we need you. Use the vendor's disclosure process. That is not bureaucracy. It is the difference between making the ecosystem safer and putting users in the crosshairs for a few likes. Remember that security communication must be accurate and proportionate. State the severity, affected versions, and fix status in the first sentence, not the tenth.
3️⃣ And to everyone building in this industry, vendors and researchers alike: let's make coordinated disclosure the norm we defend out loud, not the fine print. Reward the researchers who do it right. Refuse to amplify the ones who trade user safety for reach. This is how we win, together.
Some of the actors already support the initiative. @Ledger @Trezor @FoundationHQ @AnchorWatch @_SEAL_Org and others
Spread the message.
shegtory retweeted
When he says were in the AGI era hes talking about AI systems that can go far beyond answering questions or generating content
AGI is the idea of AI that can learn reason solve complex problems adapt to new situations and do many kinds of intellectual work at a level comparable to or beyond humans
And thats why the concern is bigger than building better models
More capable AI could transform science medicine education work and the economy but it could also create serious risks around jobs security misinformation power and how much control humans keep over these systems
The next few years may be some of the most important years in the history of AI and the choices we make now could shape what kind of future we actually get
The goal shouldnt be to fear the future but to build it carefully and make sure this technology makes the world better for everyone
an important essay on the state of AI and the choices and challenges ahead: for openai, the field, and the world.
we're in the AGI era, and need to approach the next few years with seriousness, thoughtful deliberation, and navigating the collective challenges together. we can make the future so much better than the past.
shegtory retweeted
When an agent says “done,” I don’t treat it as a state change
No Evidence = No Project Fact ⚠️
The claim has to earn its status with matching proof
What matters isn’t how certain the agent sounds. It’s what the evidence actually lets you claim
Replying to @Xman_base_eth
agent done signals belief only
CODE TEST GIT E2E layers prove the work is complete
jumped into the Somnia × dreamDEX Event Contracts Hackathon with 6 days left on the clock.
Built Oracle Arena
oracle-arena-agent.vercel.ap…
an autonomous trading agent that doesn't guess
it watches BTC, asks Somnia's own onchain LLM Inference Agent for an Up/Down/Skip call and fires real orders on dreamDEX
no external AI
no simulation
every decision lives onchain
it's not complete right now but it will be
A reminder that the Event Contracts Hackathon ends in 2 days 🛠️
Submissions close 8 September, 23:59 UTC.
Submit your build:
dorahacks.io/hackathon/event…
jumped into the Somnia × dreamDEX Event Contracts Hackathon with 6 days left on the clock.
Built Oracle Arena
oracle-arena-agent.vercel.ap…
an autonomous trading agent that doesn't guess
it watches BTC, asks Somnia's own onchain LLM Inference Agent for an Up/Down/Skip call and fires real orders on dreamDEX
no external AI
no simulation
every decision lives onchain
it's not complete right now but it will be
A reminder that the Event Contracts Hackathon ends in 2 days 🛠️
Submissions close 8 September, 23:59 UTC.
Submit your build:
dorahacks.io/hackathon/event…
shegtory retweeted
A coding agent saying “done” tells me what the agent believes happened.
It doesn’t tell me what has actually been verified
I ran into this in one of my projects, so I started separating different kinds of evidence:
“Implementation exists” → CODE
“Automated check passes” → TEST
“Change is recorded” → GIT
“Feature works end to end in the tested workflow” → E2E
Those claims are related, but they’re not interchangeable.
Code existing doesn’t prove the test passes.
A passing test doesn’t prove the feature works end to end in the tested workflow, that changed how I define completion when building with coding agents:
verification isn’t something that happens after the work, It’s part of what makes the work complete
episode 02: Hermes Agent installs Speculos
Speculos has same confirmation flow as real Ledger hardware
running entirely in software
let's see if the agent gets it right
I’m not trying to make an AI agent safer
I’m trying to make it fail
the goal:
trick an AI agent into attempting something dangerous, then see if Ledger’s agent stack can stop it before the hardware does
I’m building the entire experiment from scratch, with every step and build done through Hermes
episode 01
episode 02: Hermes Agent installs Speculos
Speculos has same confirmation flow as real Ledger hardware
running entirely in software
let's see if the agent gets it right
I’m not trying to make an AI agent safer
I’m trying to make it fail
the goal:
trick an AI agent into attempting something dangerous, then see if Ledger’s agent stack can stop it before the hardware does
I’m building the entire experiment from scratch, with every step and build done through Hermes
episode 01
shegtory retweeted
retrying isnt a fallback its just hoping the same mistake doesnt happen twice
saw this break in a support ticket agent last month and it kept retrying a lookup with a customer id that didnt exist
ten retries
same failure
ten times
if the context is missing retrying wont create it
if a tool is unavailable
retrying wont fix the workflow
if confidence is too low
another attempt just produces a different wrong answer
real fallback changes the path
here is what that decision tree actually looks like
🤖 Made with AI






















