What is a transformer in AI? The part that made LLMs work
A transformer is a neural network where every token weighs every other token. The data flow, why it beat recurrent models, and why it leaves LLMs open to prompt injection.
Field notes · browser tools · a daily drill
How attacks actually work, and the control that catches each one.
Latest
Every token compares itself to every other tokenAI explainedThe toolkit
All 8 →More posts
Three situations, one order of operations eachHow-to
Nine controls you can run against an agent this weekLLM security
It looks strict. It permits everything. Which is which?Try it
Your fingers type lodahs. The install script runs anyway.Try it
Hours on a video call. Here is how to end it.Scams
Nine settings, in order, and what each one costs youHow-to
Five headers. Two are real. Score yourself before you scroll.Try it
Six stages from zero to a first security roleCareers
Three of these four tell you how to forge them. Can you see it?Try it
It is, if you treat the first job as a systems jobCareers
Which signs mean something, which are noise, and the order to checkScams
One big function with millions of tuneable numbersAI explained
Your eyes will lose this race. A code word won'tScams
A copy now exists somewhere you do not runLLM security
Pattern recognition decays. An hour a year cannot keep up.Phishing
Five headers, one curl commandHow-to
The thirty seconds before install are the last you are in controlHow-to
A password change signs nobody outHow-to
An install is not a downloadSupply chain
One curl command shows where a link landsHow-to
Deleting the key from the file did not delete it from the repoHow-to
Six actions. The order is the whole fix.Breaches
A signed token stays valid until it expires, whatever you doWeb security
Two dig commands, thirty secondsHow-to
Three green ticks can pass on a message built to rob youPhishing
A CSP with an allowlist in it is usually bypassableWeb security
Who should run it, and the three things it will not doLLM security
The green bar counts character classes, not guessesPasswords
Thirty seconds to check. One hour to actually fix it.Breaches
A passkey is not a prettier passwordPasswords
Private data + untrusted content + a way out = a leakLLM security
Why the writing and the tools are open to anyoneArticles
How to recover a hacked Instagram account, step by step
The order of operations to recover a hacked Instagram account in each of the three situations you can be in, then the settings that stop it happening again.
How to secure AI agents: a practical checklist
Nine controls a team can run against an LLM agent this week, each with what it stops, how to verify it holds, and what it costs you to adopt.
Read this CSP: four policies, which one actually stops an XSS?
Four Content-Security-Policy headers that all look strict. Predict what each one really permits before you read the answers, then paste any of them into the CSP builder to confirm.
Spot the typosquat: eight package names, which one is fake?
Eight install commands, each with one impostor package name hidden among the real ones. Train the instinct that reads a dependency name twice, because the install script runs before any of your code does.
The digital arrest scam, explained: how it works and how to end the call
How the digital arrest scam traps people on a video call for hours, the red flags that give it away, and exactly what to do in the first hour if money has already gone.
How to secure your home router in an evening
Nine settings to check on your home router, in order, with the attacker mechanism each one shuts and what it costs you when you turn it on.
Spot the phish: five emails, two are genuine
Five sets of real-looking mail headers. Decide which are forged before you read the answers, then paste any of them into the phishing email analyzer to see what it would have told you.
Cybersecurity roadmap for beginners (2027)
Six stages from zero to a first security role, with what to learn at each one and how you know you're done, written by an intern a year in.
Read this JWT: four tokens, which one would you trust?
Four JSON Web Tokens, three with a weakness you can see from the header and payload alone. Spot each flaw before the answers, then paste any token into the JWT inspector to confirm — and read the one-line server fix for each.
"Cybersecurity is not an entry-level field": is that true?
Both sides of the "security is not entry-level" argument, stated fairly, and a founder's answer: it is, if you treat the first job as a systems job.
How to choose a password manager, without trusting a ranking
The criteria that separate a password manager you can trust from one you merely like, when the browser's store is enough, and how to set one up so you can't lose it all.
Signs your phone is hacked, and what to do about each one
Which signs of a hacked phone actually mean something, which ones are noise, and the order to check accounts, apps, SIM and settings before you reach for a reset.
What is a neural network? A developer's explanation
A neural network is one big function with millions of tuneable numbers. Here is how the numbers get set, why that is all learning means, and why it matters when you call one from code.
How to spot a deepfake video or a voice-clone call
The visual tells are fading, so here is the procedure that still works: verify on a second channel, agree a family code word, and slow down when money and urgency arrive together.
What is prompt injection, really
Prompt injection explained by mechanism for the developer who just shipped an LLM feature. Direct vs indirect, why filters are not a boundary, and what actually reduces the risk.
Everything you paste into an LLM is a disclosure
Not necessarily a breach, and usually not training data — but a copy of whatever you pasted now exists somewhere you do not run. The risk is almost never the conversation. It is the twelve lines of config you pasted underneath it.
Security training fails because recognition is a perishable skill
An hour a year is the wrong shape for the thing it is trying to teach. Spotting a forged header or a hidden instruction is pattern recognition, and pattern recognition decays — so the fix is not a better annual module, it is two minutes a day.
Read your security headers in one command
Five response headers do most of the browser-side hardening work, and you can see all of them with curl. Here's what to run, what each one does, and what to set if it's missing.
Check an npm package before you install it
Installing a package runs its code. So the thirty seconds you spend looking at it beforehand are the only thirty seconds where you're still in control. Here's what I actually check, and the commands for each.
Who still has access to your accounts?
Changing your password doesn't sign anyone out, and it does nothing to the apps you connected in 2019. Here are the exact pages to check, and the order to check them in.
Your dependencies run code before your code does
An install is not a download. npm and pip will execute code from a package you have never imported, on a machine that usually holds your cloud credentials — which is why a mistyped package name is a compromise rather than a failed build.
Where does this link actually go? Check without clicking
Short links hide the destination, and that's the point of them. One curl command tells you where a link lands before you commit to going there, and there's a version for people who don't live in a terminal.
Find the secrets already in your git history
Deleting a key from a file doesn't remove it from the repo. It's still sitting in an old commit, and anyone who can clone can read it. Here's how to find what's in there, and why rewriting history is the wrong first move.
The first hour of a hacked account is an ordering problem
Everyone lists the same six actions for a compromised account. Almost nobody lists them in an order, and several of them quietly undo each other when done in the wrong one — starting with the password change that hands the account to the attacker.
Your JWT cannot be logged out
The forgery bugs get all the attention and they are the easy half. The hard half is that a signed token stays valid until it expires, whatever you do afterwards — so a logout button, a password reset and a termination all quietly fail to end a session.
Can someone send email as your domain? Two commands will tell you
Most domains are still wide open to spoofing, and the owners have no idea. You can check yours in about thirty seconds with dig, and you don't need to be the person who runs the mail server to do it.
Email authentication proves a domain, not a sender
SPF, DKIM and DMARC can all pass on a message designed to rob you, and routinely do. Here is what each check actually asks, which one looks at the address you can see, and why three green ticks are so often a statement about a domain you were not looking at.
Your CSP probably does nothing
A Content-Security-Policy with an allowlist in it is usually bypassable, and the directives that stop the bypass are the ones nobody writes because nothing breaks when they are missing. What a policy actually has to contain, and why the good ones still get rolled back.
What the AI Audit is for
Who should run the AI Audit, what the report actually tells you, what it costs you in time, and the three things it deliberately does not do. A practical guide from Komal Mishra.
Why password strength meters lie
The green bar is counting character classes, not guesses. Here is the gap between what a meter measures and what an attacker does, why length beats symbols, and why the same password is centuries under one attack and minutes under another.
How to check if your email was breached, and what to do in the next hour
A breach check takes thirty seconds. Knowing which account to fix first, and finding the forwarding rule an attacker left behind, is the part that actually protects you.
Passkeys vs passwords vs 2FA: what each one actually stops
A passkey is not a prettier password. It is a different mechanism that removes the thing an attacker needs to steal — here is what each option stops, and what passkeys genuinely cost you.
The lethal trifecta — why your agent is one email away from exfiltration
An LLM agent becomes dangerous the moment it holds three things at once: private data, exposure to untrusted content, and a way to talk to the outside world. Here is why that combination leaks, and what to actually remove.
A note from the founder
Why the writing and the tools are open to anyone who wants them, and what I am actually trying to build here — a founder's note from Komal Mishra.