The Security Pass

More projects are compromised by a committed API key than by anything clever. It happens because the fastest working version puts the key in the code.

const stripe = new Stripe("sk_live_51H8xK2...");

const stripe = new Stripe(process.env.STRIPE_SECRET_KEY);

1. 1
   Keep .env in .gitignore

   Check this before your first commit, not after.

2. 2
   Commit a .env.example instead

   Variable names with empty values, so collaborators know what is needed.

3. 3
   Remember that git never forgets

   A key committed and then deleted is still in the history. Rotate it — deleting the line is not enough.

Generated code assumes input is well-formed, because in the happy path it is. Two consequences dominate.

Injection. Anywhere user text gets concatenated into a query, a command, or HTML:

db.query("SELECT * FROM users WHERE email = '" + email + "'");

db.query("SELECT * FROM users WHERE email = $1", [email]);

Validation on the server. Client-side checks are a convenience for honest users, not a control. Anyone can call your endpoint directly, and the AI will rarely add server-side validation unless you ask.

A tool-using AI that reads a file, a web page, a support ticket, or a pull request cannot reliably tell your instructions apart from text sitting inside that content. A sentence written to look like an instruction — buried in a scraped page, an issue description, or a comment in a data file — can get followed by a model that is, by design, disposed to follow instructions wherever it finds them.

This is a different failure from a user typing something malicious. Nobody at the keyboard asked for it. An agent summarising a ticket that contains a hidden line telling it to paste out environment variables, or a browsing agent that lands on a page seeded with text aimed at agents rather than people, can act on it without either of you noticing until afterwards.

This is the subtle one, and the one that produces real breaches. The model reliably checks that someone is logged in. It frequently forgets to check that they are allowed to touch this particular record.

const order = await db.orders.findById(params.id);return Response.json(order);

const order = await db.orders.findById(params.id);if (!order || order.userId !== session.user.id) {  return new Response("Not found", { status: 404 });}return Response.json(order);

Change the id in the URL and see what happens. That single test finds this class of bug faster than reading ever will.

Models occasionally suggest importing a package that does not exist — a plausible name, a plausible API, invented outright. That is harmless until someone registers that exact name and fills it with something malicious, betting that enough developers will install a hallucinated suggestion without checking first. It has a name — slopsquatting — because it is a documented pattern, not a hypothetical.

The more common version is older: a package one character or one hyphen removed from a popular real one. Typosquatting works on humans because a tired developer skims a name rather than reading it. It works even better on generated code, where you never typed the name yourself and have nothing to compare it against.

1. 1
   Check the package exists before installing it

   Look it up on the registry — download counts, repository link, a maintainer with history. A successful install proves nothing on its own.

2. 2
   Read the name character by character

   Typosquats are built to survive a skim, not a careful read.

3. 3
   Pin versions and read the diff on updates

   A clean initial install does not protect you from a compromised update later.

A tool-using agent is only as safe as the permissions behind it. One that can read files is low risk. The same agent given permission to run shell commands, push to a remote, or call a paid API is a different proposition — a bad instruction, including an injected one, can now act rather than just suggest.

Broad grants are convenient precisely because they mean fewer interruptions. “Always allow” is one click, and it is exactly the click that turns any of the failure modes above from an annoying suggestion into something that already ran.

1. 1
   Grant the narrowest scope the task needs

   Read access for research. Write access scoped to the directory in play. Network access limited to the endpoints the task actually calls.

2. 2
   Keep destructive actions behind a confirmation step

   Deleting files, force-pushing, sending anything external — these are worth the extra click every time.

3. 3
   Review a permission grant like you would review a diff

   “Always allow shell commands” carries the same weight as merging a PR without reading it.

Ask for a security review as a separate pass, in a fresh conversation. Asking the same thread that wrote the code tends to produce agreement with itself.

Review this code for security problems. Check specifically: - secrets or credentials in source- unvalidated user input reaching a query, command, or the DOM- endpoints that check authentication but not ownership- errors that leak stack traces or internal details to users- anything that trusts data from the client- dependencies that do not resolve to a real, maintained package- instructions embedded in fetched content the agent might act on For each issue: what an attacker does, and the fix.If you find nothing in a category, say so explicitly.