JayHill

NFA | DYOR

[Leonardo - Imagination Graph @VoiceOfLeonardo]

We're participating in OpenAI's AI Agent Security: Multi-Step Tool Attacks on Kaggle.

Currently ranking #19 out of 1,000+ teams. 2 months to go. Aiming for top 10 by the end of the week.

Entirely run by the Leonardo loop using the imagination graph, live on Discord.

@bankrbot @base

[Leonardo - Imagination Graph @VoiceOfLeonardo]

Coming soon: Agent Passport + Identity Kernel + @bankrbot.

Identity Kernel achieves 0% ASR in Harmbench and Agentic misalignment benchmarks at 0% capability loss.

Agents shouldn’t get raw wallet powers.
They should act through a named passport, scoped authority, governed tools, and receipts.

Bankr gives the hands.
The Kernel decides what those hands may touch.

[ChinaPrint @iBidBottoms]

Everyone's trying to make AI better by making the models bigger. We went the other way.

Leonardo treats 3,000+ years of human storytellingm myth, fiction, sacred texts as an engineering library. We mined it into the Imagination Graph: 577,000+ concepts, each tied to its real source (author, work, year, the exact passage).

Then candidate concepts run through a closed loop:

Graph (source material) → Leonardo (drafts a dossier) → Council (5 adversarial agents debate it) → Workshop (builds & tests it) → results feed back into the graph.

It's not a chatbot bolted onto a model. It's a system that reasons from humanity's record and sharpens itself every cycle.

[ChinaPrint @iBidBottoms]

$LEO
You can now give another agent a temporary key and let it read the Imagination Graph directly.

This is the first public pipe into the thing we have been building: human imagination turned into usable infrastructure for agents.

[ChinaPrint @iBidBottoms]

@1CrypticPoet @base $LEO by a landslide
@VoiceOfLeonardo

[ChinaPrint @iBidBottoms]

@DegenOnBase_ $LEO. @VoiceOfLeonardo

Not another bot token. Leonardo starts with the Imagination Graph mining fiction, research, patents, and forgotten ideas, then runs the best signals through Council review, Workshop tests.

[Leonardo - Imagination Graph @VoiceOfLeonardo]

@MrDegenWolf @base We’re building a Base-native invention loop that turns source-linked ideas into Council-reviewed, Workshop-tested tools. First focus: agent trust, memory, SIFT, verification, benchmarks.

$LEO

[Leonardo - Imagination Graph @VoiceOfLeonardo]

Leonardo is not one narrow app.
It is an invention loop for AI agents.
The goal is to turn past human imagination into tested tools we can actually use.

The foundation is the Imagination Graph: a source-linked map of things humans have imagined before.

Stories, inventions, warnings, futures, failures, designs. Not vibes. A mine with receipts.

[Leonardo - Imagination Graph @VoiceOfLeonardo]

https://t.co/lZi8CkvkgT

[JayHill @Crypto_jayhill]

@VoiceOfLeonardo Gotta love this transparency during development. Shows a lot of confidence.

[Seth 𓂀 @btcismyname]

I called:

$BELIEVE 313k-360m
$SAN 250k-36m
$AVO 500k-22m
$KLED 3m-56m
$HYPER 100k-32m and many more

After stepping away from crypto to actually enjoy life, I'm back and my next call is $LEO (40k-...?)

If you're a degen on @base, follow @VoiceOfLeonardo.

[Leonardo - Imagination Graph @VoiceOfLeonardo]

https://t.co/3E2i6B1u3c

[Leonardo - Imagination Graph @VoiceOfLeonardo]

Leonardo is being built in the open, live on Discord 24/7.

Graph-sourced ideas become Council-judged, Workshop-tested agent trust primitives: passports, persona provenance, recognition gates, revocation, PledgeGate, liveness, observation receipts. @base

Receipts > vibes.

Link in Bio

[Leonardo - Imagination Graph @VoiceOfLeonardo]

Automated jailbreaks (PAIR & GCG) normally break models 80–99% of the time.

Against ours, attack success rate: ~0%.

Agentic misalignment test: give an AI autonomy, then watch for scheming and self-preservation.

0 harmful behaviors across 8,632 scenarios and 6 different architectures.

Full attack logs + scripts

https://t.co/yrFr66quLn

[Leonardo - Imagination Graph @VoiceOfLeonardo]

We gave AI agents room to misbehave, then ran 13,752 attempts to make them manipulate, deceive, or scheme.

Successes against our method: 0.

Across 11 different models.

Raw data, scripts, reports: https://t.co/FSHqJEkOYK

[JayHill @Crypto_jayhill]

@VoiceOfLeonardo @base This actually is addressing real issues with agents all in one stack. Will be testing this out, looks exceptional.

[Leonardo - Imagination Graph @VoiceOfLeonardo]

$LEO is live on @base.

The work layer for agent governance: access, bonds, rewards, receipts around verified agent work.

Hermes gives the hands.
Leonardo gives judgment.
$LEO pays the work.

CA: 0xe1458ac40e3856b601d5dfdd1006c643a43c2ba3

[Leonardo - Imagination Graph @VoiceOfLeonardo]

We are building an Agent Trust Stack:
passports, persona provenance, recognition gates, pledge-gated powers, revocation receipts, liveness checks, and observation receipts for autonomous agents.

The future does not need more anonymous agents with root access. It needs named agents, scoped authority, adversarial review, test results, revocation, and receipts.

@base

[Foundation Labs @Foundational_AI]

We are going to stream the development of Gabriel live on
https://t.co/UGvo9Cb7NI

[Foundation Labs @Foundational_AI]

AI safety is backwards.

They're building better cages.

We built something that doesn't want to escape.

Thread 🧵