~/blog/labs
AI and ZK Security Labs
Experimental learning notes on AI security, machine-learning attack classes, zero-knowledge proofs, and adjacent security research.
AI Security ML10: Model Poisoning
Rewriting the brain from the inside. Goes straight to the model's parameters — the numerical weights that encode everything it has learned — and edits them to produce attacker-controlled behaviour.
read moreAI Security ML09: Output Integrity Attack
The model was right. The answer was wrong. Every other attack tampers with the model — this one intercepts the correct answer on the way out and replaces it with a lie.
read moreAI Security ML08: Model Skewing
Slow poison through the feedback loop. The model wasn't broken on day one — it was steered slowly through fake feedback until its decisions aligned with exactly what the attacker needed.
read moreAI Security ML07: Transfer Learning Attack
The backdoor you inherited. The model passed every benchmark and scored well on your validation set — but it was trained to fail in exactly the one scenario the attacker controls.
read moreAI Security ML06: ML Supply Chain Attacks
The poison was already in the package. Did you audit every library your training script imported? Every model you downloaded from a public hub? The attack was already upstream.
read moreAI Security ML05: Model Theft
Your API is the heist. If a model is queryable, it is reconstructable. Every input-output pair is a brick — and a patient attacker will collect enough bricks to build your model themselves.
read moreAI Security ML04: Membership Inference Attack
You don't need to extract the training data. You just need to ask the right questions — and the model will tell you whose data it learned from, one confidence score at a time.
read moreAI Security ML03: Model Inversion Attack
Reconstructing training secrets from nothing but model outputs — how attackers invert a model's predictions to recover the sensitive data it learned from.
read moreAI Security ML02: Data Poisoning Attack
A handful of poisoned training samples can flip a model's behaviour on exactly the inputs an attacker cares about — here's how the attack works and where to defend.
read moreAI Security ML01: Input Manipulation Attack
Adversarial inputs that look normal to humans but flip a model's decision — the attack class that breaks fraud filters, content moderation, and on-chain ML oracles.
read moreZK Journey Week 5 — Learning Zero Knowledge Proofs from Scratch
As a Move security researcher, I need to understand ZK proofs deeply — not just what they do, but how they work under the hood. Week 5 of my ZK learning journey at the Rare Skills ZK Bootcamp.
read moreZK Journey Week 4 — Learning Zero Knowledge Proofs from Scratch
Week 4 of my ZK learning journey at the Rare Skills ZK Bootcamp.
read moreThe Move Auditor
Introducing move-auditor — an open-source Claude skill that runs a structured security-audit pipeline over Move codebases, built from my real engagement workflow.
read moreZK Journey Week 3 — Learning Zero Knowledge Proofs from Scratch
Week 3 of my ZK learning journey at the Rare Skills ZK Bootcamp.
read moreI Open-Sourced My AI Audit Workflow — Prompts, Pipelines, and Everything I Learned
The full prompt pipeline I use to assist real audits — recon, hypothesis generation, verification gates — published with notes on what worked and what didn't.
read moreZK Journey Week 2 — Learning Zero Knowledge Proofs from Scratch
Week 2 of my ZK learning journey at the Rare Skills ZK Bootcamp.
read moreZK Journey Week 1 — Why a Move Security Researcher Is Learning Zero Knowledge Proofs from Scratch
As a Move security researcher, I need to understand ZK proofs deeply — not just what they do, but how they work under the hood. Week 1 of my ZK learning journey at the Rare Skills ZK Bootcamp.
read more