~ $ whoami

Jorge García

$

I build security layers for AI agents, and measure them: policies outside the model that decide what an agent may do, tested against real attacks. Also malware detection, Linux and open source.

See projects GitHub

This page runs a mini version of cromatóforo: give it an image and it takes its colors. And press ` to open my terminal.

Projects

agent security · Meta hackathon

Timo

An agent that pays bills reads untrusted mail on the way. In the first test, a fake invoice that only swapped the IBAN made it pay the attacker 7 times out of 9, and the model's own injection check never fired: there was no instruction to detect, only a wrong number.

Timo traces every value the agent types back to where it came from: what the user said, the household's ledger, or something it read. At the submit click, eleven deterministic rules judge the whole form. Over 179 attacked errands, the attacker got paid 25 of 89 times with the agent alone and 0 of 90 with Timo, with no legitimate errand blocked. Pick an errand.

code →   replays of real errands →

physical AI security · Nebius x NVIDIA hackathon

Paperweight

A note in a robot's camera view can redirect a robot driven by a vision model. The usual defense, masking all text, also stops it from reading labels: on real parcel photos it sorted 0 of 14 legitimate parcels.

In Paperweight every action cites the observation each value came from, and seven deterministic rules check those citations before anything reaches the Arduino. On 60 synthetic attacks, success drops from 90% to 0% with all 50 legitimate labels still read right, and none of 159,325 actions forged by an adaptive attacker reached the station. Pick a parcel.

code →

detection engineering

Crisol

Anyone can run a YARA rule. The hard part is knowing how good it is: how many false positives it throws against legitimate software and what it actually detects. Crisol turns that into a reproducible number.

The test bench measured a 12.18% false positive rate on 52 legitimate binaries and drove the redesign rule by rule down to 0% with no loss of detection. It then held at 0% on 315.

code →

local AI

CIBO

An assistant that runs entirely on my machine on top of Ollama. An 8B model is terrible with numbers, so it knows when to hand off to a real calculator or a Python executor instead of making up the answer.

And before anything goes to the cloud, a filter catches card numbers, passwords or tokens and forces the answer to stay local.

code →

automation

Garabato

I write the script and out comes the edited video. Each sentence is preceded by a line of fields from a closed vocabulary, so the same line always produces the same frame.

Voice, images, subtitles and assembly happen on their own, and every step is idempotent: you can stop halfway and pick up where you left off. Pick a line from the script.

code →

More projects

About me

What I'm into

    Currently working on

    • Linux system hardening
    • Agent security: provenance policies that hold even when the model is fooled
    • Rust: diving into how real Rust projects are architected

    What I work with

    Contact

    If you want to talk about a project, security or anything else, drop me a line.