Welcome to Safety in the Layers
What this site is, the 180-day plan behind it, and the four kinds of posts you'll find here.
Illustration: a language model completing “safety in the …”. The next-token dropdown chooses “layers” (0.91) and strikes out “jailbreak” (0.06) and “breach” (0.03).
Learning AI safety & security in public
I'm Prathik, a security engineer spending the next six months locked in on alignment, interpretability and adversarial robustness. Everything goes here: the daily log, the papers, the builds, and the exact roadmap. Follow along, or take the same path.
Day 7 of 180 · Week 2: Projections, eigen-structure, SVD, calculus, Karpathy
The lock-in
180 days, logged in public.
Each week of the plan is one token in the context window. The glowing one is where I am right now.
7day of 180
2week of 26
3.9%complete
Today on the plan
Week 2 · Projections, eigen-structure, SVD, calculus, Karpathy
- MStrang lectures 14–16 (projections, least squares); implement the projection matrix in NumPy
- CModule 7: Attacking AI, Application and System
- BRepo hygiene: config files, seed handling, reproduce.sh stub
- RA Mathematical Framework for Transformer Circuits, part 1
- LNotebook entry, commit
What I post
Four kinds of notes.
The same work, cut four ways: the day-by-day log, the papers behind it, what I build, and what I can teach back.
The Log
Daily and weekly notes from the lock-in: what I studied, built, broke and learned.
1 postPaper Notes
Research papers broken down: the claim, the method, the numbers, and what I think.
First post soonBuilds & Research
Things I'm building and measuring: harnesses, attacks, evals, and the results.
First post soonExplainers
Concepts explained the way I wish they had been explained to me.
First post soonLatest
From the notebook
Logs, paper notes, builds and explainers, newest first.
All postsProjects
What I'm building and measuring
Attack Headroom
When gradient-based adversarial attacks fail against safety-tuned language models, how much of that failure is real robustness, and how much is the search, the harness, or the grader?
Where refusals actually come from
How does the framing of a legitimate security request move a model's refusal boundary?
Local LLM evaluation rig
Which open-weight models are actually useful for security work on a single 16 GB GPU, and how do you measure that honestly?
New here?
The path I'd hand a friend who wants into AI safety.
Pick a lane, then build foundations, hands-on skills, papers, projects, programs and community, with the free resources that actually matter at each step.