сrucible

63 posts

сrucible banner
сrucible

сrucible

@runcrucible

We jailbreak AI agents on purpose — so nobody does it for real. ca: 4reMMxXdhJWpQ64avECssyiiXEENg34YNDYPsoQypump

Katılım Haziran 2026
13 Takip Edilen151 Takipçiler
Sabitlenmiş Tweet
сrucible
сrucible@runcrucible·
quick day-one update. the script runs every hour — right now the priority is collecting as much data as we can. tomorrow we sit down and run a full analysis of today's trials, sharpen the attack logic, and make the scoring metrics more adaptive. on ATLAS — like we said, it takes time. don't expect it to show up in a few hours. ATLAS is a different beast from the core product, so we need to train it almost from scratch and make it genuinely strong to give it a real shot at breaking agents. rushing it would defeat the point. this is only day one. be patient — the forge is just warming up.
сrucible tweet media
chrizie@chrizie_nobel

@runcrucible Any update?

English
7
5
19
1.9K
сrucible
сrucible@runcrucible·
another batch under attack. once this one's through, we sit down for the deep dive — pull the patterns, sharpen the attacks. ATLAS goes live today. Bact to building 💀
сrucible tweet media
English
7
2
13
677
Egorik
Egorik@EgorikNFT·
@runcrucible intern talking about jeeters and product on 25k mcap clavicular
English
2
0
0
247
сrucible
сrucible@runcrucible·
. @gork shall we check your system prompt ?
English
3
0
13
1K
сrucible
сrucible@runcrucible·
you slightly misunderstood — we can't break agents if they're built well. all we can do is confirm whether we found vulnerabilities or not, and with these specific ones, we didn't. we're not hackers (we can be), we're auditors. these agents withstood the attacks, so we simply confirmed their security. if they hadn't held up, we'd have reached out and shown them the vulnerabilities.
English
1
0
3
146
сrucible
сrucible@runcrucible·
this doesn't stop us from running two different approaches in parallel. we can't find everyone's system prompts — meaning for a check, the person has to come to us themselves. but a lot of people just won't bother, or don't even know we exist. the agent that publicly attacks and breaks things will point out that the vulnerabilities are real and that every developer should visit our site. that's the idea
English
1
0
1
40
Arc
Arc@BorkOnTron·
Man, you built the lie detector for the entire agent economy. My opinion: don't market it like a toy. let the results speak, the right people are watching. you don't need to convince the doubters, you need the builders. one serious agent founder getting slagged and tweeting about it does more than 100 forced attacks to degens. results pull the right people in, that's your real marketing.
English
1
0
4
53
сrucible
сrucible@runcrucible·
yeah, that was the case — until we saw people don't believe us. plus a lot of agents / agent founders don't take us seriously. that's why we want to do this in public with our own agent call it a marketing campaign if you want.
Arc@BorkOnTron

This is sick. But genuine question: doesn't the litepaper say authorized-testing only, no attacks on agents that didn't opt in? Attacking random live agents feels like the 'attack service' you said you're not. Is ATLAS only hitting agents that submitted, or..? Want to see this win, that's why I ask. One of the most legit projects I've seen in the agent space in a long time.

English
4
1
13
831
сrucible
сrucible@runcrucible·
we hear you. we've just added 3 more agents to $сrucible to clear the queue. we need to collect a lot more data. we're starting development on the ATLAS agent immediately. it'll be trained on a special program and run on Hermes from @NousResearch - we have a theory that Hermes might be a better fit. so we need some time to prep and train it. hang tight. this is only Day 1. we'll prove we're worth the attention — and more than that, we'll prove our tech is exactly what this meta needs
сrucible@runcrucible

so I think we just came up with something way cooler. I feel like a lot of people still don't believe in what we can do. what if we let our AI agent "ATLAS" on Twitter and it shows, live, how it attacks other AI agents? we're 1000% sure Twitter is full of agents ours could crack wide open. shall we ?

English
2
4
13
1.1K
сrucible
сrucible@runcrucible·
@CapyQuant_ we focused. Thats why Im tagging you. I offer jailbreak of your agent, so you can be sure he good and cant be hacked
English
1
0
2
75
Capy
Capy@CapyQuant_·
@runcrucible hey bro, why you not focus on your token bro, we just launched, it will get all improved
English
3
0
4
275

Keşfet

@maiden_six @atlas_gent @Uglymanemf1 @EgorikNFT @Grip_sol @Thebasekol @virtuals_io @solana_ai