Amy Lily

Jailbreaking LLMs: When the Chef Decides to Go Rogue

Wed, 29 Apr 2026 00:00:00 +0000

Prompt injection tricks the chef with a fake note. Jailbreaking convinces the chef to want to break the rules. Here’s the theory, the research, and the tools that show why no model is fully immune.

From Infra Chaos to a Personal AI Operator: Building with OpenClaw

Tue, 07 Apr 2026 00:00:00 +0000

TL;DR

Built a personal AI operator using OpenClaw + Telegram + Claude. Got stuck optimizing too early (local models, networking). Reset everything. Used OpenClaw onboarding + Anthropic API. Connected Telegram. System worked instantly.

Biggest lesson: Getting something working > getting everything perfect.

Introduction

Over the past few days, I set out to build something simple:

a personal AI operator I could interact with through Telegram.

What I ended up building was not just a chatbot, but a small system:

Same Model, Different Answer: What System Prompts Actually Do

Mon, 06 Apr 2026 00:00:00 +0000

I asked three Claude interfaces the same prompt injection question. One gave me a full strategy guide. The other two drew the line at explanation. Same model — completely different behavior. Here’s what that reveals.

Prompt Injection: Hijack LLM Instructions

Thu, 26 Mar 2026 00:00:00 +0000

I started a Hack The Box AI security lab with zero expectations. Three minutes later, I had leaked a secret key the model had been explicitly told never to reveal. Here’s what I learned — and why it matters if you build with LLMs.

AIBadge: A Weekend Experiment about AI Provenance

Mon, 09 Feb 2026 00:00:00 +0000

summary: A meeting question about AI disclosure led to a weekend experiment exploring trust, authorship, and verifiable integrity in the age of generative AI.

From Syntax to Semantics: Prototyping a Natural Language Interface for Secure Infrastructure Search

Thu, 12 Jun 2025 17:15:51 +0100

Context:

MAAS search historically relied on a custom Domain Specific Language (DSL) — precise but intimidating and error-prone for both expert and new users.
The translation step (from human intent to DSL syntax) was the biggest source of mistakes, rework, and delays.
Users frequently forgot query syntax between sessions, making search a recurring learning curve.

Challenge:

How might we remove the manual translation step while keeping search fast, precise, and scalable?

whoami

Sun, 01 Dec 2024 16:20:00 +0000

👋 Hi, I’m Amy!

UX explorer by day, dumpling enthusiast by night 🥟✨

I’m a London-based UX Design Manager with a soft spot for elegant systems, bold ideas, and anything that makes complex things feel delightfully simple. I’ve spent the last 10+ years designing products and experiences that live at the intersection of cloud infrastructure, AI, and open source — making the technical feel human.

My favorite things include:

Ubuntu Pro CLI: Redesigning `pro status` for Clarity, Scale, and Actionability

Fri, 12 Apr 2024 17:15:51 +0100

Context: Ubuntu Pro provides security, compliance, and extended maintenance services for enterprise and developer environments. The pro status CLI is the primary entry point for human operators to check subscription status, security coverage, vulnerabilities, and service health.

As the service portfolio grew across LTS, EOSS, non-LTS, and cloud-attached systems, outputs became inconsistent, hard to parse, and difficult to act on — especially for operators managing dozens or hundreds of machines.

From Confusion to Clarity: Designing `/explain` to Make Concepts Click

Sat, 12 Mar 2022 17:15:51 +0100

Context:

At Canonical, internal teams often work in highly specialized domains — cloud infrastructure, AI, enterprise support, etc. As a product designer, I frequently saw confusion arise when teams shared updates full of acronyms, assumptions, and product-specific jargon.

Why?

During a design review, I realized that while I thought I had clearly explained the mental model behind LXD integration in MAAS, my colleagues didn’t follow.

Little did I know, this was how other designers understood my explanation.

Scaling Infra UX: Designing a Centralized Control Platform for 60K MAAS Sites

Sat, 12 Mar 2022 17:15:51 +0100

Context: MAAS is Canonical’s bare metal provisioning tool used by enterprise and telco clients to manage thousands of machines.
As deployments scaled to 60,000+ edge sites across multi-region infrastructures, operators struggled with:

Fragmented management across thousands of MAAS instances
Manual, redundant image handling
Delayed observability and no single source of truth
Inconsistent UX between sites

I led the design of MAAS Site Manager — a centralized control platform to unify monitoring, troubleshooting, and configuration at scale.

Redesigning MAAS: Scalable Infrastructure UI That Inspired a Design System Shift

Fri, 01 Oct 2021 14:53:41 +0000

Context:
MAAS 2.8’s UI struggled to scale. Navigation was unintuitive, actions were buried far from context, and real-time status was hard to see. We set out to modernize MAAS and align the layout across Canonical’s Cloud & Infra products.

Challenge:
How might we create a functional, scalable application layout that:

Improves learnability, efficiency, and error tolerance
Scales to large environments (100–1800+ machines)
Serves as a foundation for the division’s design system

My Role:
UX Lead partnering with Engineering & PM.

The new MAAS CLI that changes Ubuntu CLI Guideline

Mon, 28 Sep 2020 17:15:51 +0100

Context:

MAAS (Metal-as-a-service) is Canonical’s bare-metal provisioning tool. The CLI was originally created for machines to interface with machines via the Foundations Cloud Engine (FCE) making it difficult for human to use. Over time, users either abandoned the CLI or built their own wrappers — indicating a breakdown in usability and trust. We set out to reimagine what a CLI could feel like when designed with empathy, efficiency, and accessibility in mind.