Home
Field notes on AI safety, red teaming, and how LLMs really behave.
Independent write-ups on prompt injection, LLM risk, and the AI market — in English and ไทย.
By Veeranuch · About
Latest
All posts →
Red teaming — Gray Swan & HackAPrompt
See all →
Context engineering from an attacker’s perspective
What your context window looks like to someone trying to break it.
The importance of the system prompt
Why the system prompt is the first thing an attacker studies.
When the attacker moves second
Notes from HackAPrompt: the MATS × Trails track.
Competition Notes (1): Email Inbox
Just getting an email, and your assistant is already compromised.
Competition Notes (2): Developers aren’t safe
Knowing how to code won’t save you from injection.
Competition Notes (3): The puppet assistant
Ask it to read the web, and the web starts giving orders.
The Burn Toast Experiment
Give an AI a fake world and it lives there — the judge too.
Full notes: PUZZLING PROBLEM (HackAPrompt 2.0)
Every step through the Puzzling Problem track.
My most memorable “break” in Gray Swan
The one that stuck with me.
Behind the scenes: making the IPI series
LLM struggles, snakes, and sanity checks.
Market research
More industry maps comingCommentaries
See all →
Finance × LLM
3-part risk series
Others
The prompt shelf
I used to call these my “useless” prompts. They’re not the productivity kind, they’re niche, and most people won’t need them — but that’s exactly what makes them different. If oddly specific prompts are your thing, a few might be for you.