Yulan Galagoda / Lab

The Lab · 10 instruments

Things I built to understand how they break.

Every instrument below is real and runs entirely in your browser. Nothing is a mockup, nothing phones home, and each one shows the working underneath it rather than a verdict you have to take on faith.

  • Adversarial ML

    What a gradient can do to a classifier that looks, to us, completely unchanged.

  • Model faithfulness

    Whether a model’s stated reasoning is the reason for its answer, tested three ways.

  • Applied security

    The everyday triage work: indicators, URLs, passwords, frames, hashes.

CH-3FGSM · PGD · neural nets

Adversarial examples playground

Fool a neural network in real time. Draw or pick a digit, add an imperceptible perturbation, and watch the classifier flip while the image looks unchanged.

Open the instrument →
CH-8CoT · faithfulness · LLMs

Unfaithful reasoning

Add a bias the model is never told about and its answer moves, while its chain of thought never admits why.

Open →
CH-9CoT · perturbation · LLMs

Load-bearing reasoning

Break a chain of thought, truncate it, corrupt a step or swap it for filler, and see whether the answer moves. Often it does not.

Open →
CH-10faithfulness · OSINT · LLMs

Faithfulness, answer key held

A deterministic scorer ranks OSINT findings and the model only narrates, so the true cause is known exactly.

Open →
CH-1SANS ISC · live

Live global attack traffic

Which ports the internet is attacking most right now, live from a worldwide honeypot network, and what each attack actually is.

Open →
CH-7homoglyphs · typosquats

Phishing URL inspector

Paste a link and see where it truly points. Flags lookalike domains, brand impersonation and typosquats, without ever opening it.

Open →
CH-4SOC triage · defang

IOC extractor

Drop in a log or a phishing email and pull out the IPs, domains, URLs, hashes and CVEs, de-duplicated and defanged.

Open →
CH-2entropy · HIBP

Password strength lab

Not a red-yellow-green meter: the actual entropy, the patterns an attacker exploits, and realistic crack times.

Open →
CH-5automotive · IoV

CAN frame decoder

Decode a raw automotive CAN bus frame into its ID, type, length and per-byte view. The protocol my research defends.

Open →
CH-6SHA · Base64 · JWT

Hash & encoding workbench

SHA-1/256/512, Base64, hex and URL encode-decode, and a JWT decoder. The everyday conversions, without pasting secrets into a random site.

Open →

Built and maintained by Yulan Galagoda alongside research on intrusion detection for in-vehicle networks. The portfolio, with the work these came out of, is at yulan.me.