2 of 3 bugs a 5KB corpus caught were the exact shapes AI writes by default

Two of the three bugs a 5KB corpus caught the first time we ran it were on the exact shapes an AI assistant emits by default — and the security rules meant to catch AI mistakes were the ones silently waving them through.

Three of our flagship ESLint rules had green unit tests for months and had been benchmarked against peer plugins on real OSS for weeks. Then that corpus — 12 fixtures, runs in 3 seconds — failed all three the first time we ran it. None of the prior signals had flinched.

The reason that should worry you before you read a single fix: each rule's widest blind spot sat exactly where AI-generated code is densest. These are security rules teams adopt to backstop AI output — and they were silently trusting the one input distribution they were never tested against. The three walkthroughs below are how that happens, line by line.

Here's how that gate works. We added a npm run ilb:flagship:smoke step to the quality script. It's small: for each flagship rule with a labeled corpus, run the rule against vulnerable/* (must fire) and safe/* (must stay silent). Compute precision, recall, F1. Fail the build below F1=1.00.

The first run hit nine rules. Six passed. Three failed.

All three rules had passing unit-test suites. All three had been benchmarked alongside peer plugins on real OSS for weeks. None of those signals would have surfaced these bugs.

What did surface them: 12 fixtures across 3 corpora — 4 per corpus, 12 to 18 lines each — labeled with // This MUST be detected or // This MUST NOT fire comments and run through the same lint config a real user would have.

Bug #1: hooks-exhaustive-deps fires on inner-callback parameters

The fixture:

import { useEffect, useState } from "react";

export function Profile({ userId }: { userId: string }) {
  const [data, setData] = useState(null);
  useEffect(() => {
    fetch(`/api/users/${userId}`)
      .then((r) => r.json())
      .then(setData);
  }, [userId]);
  return <div>{JSON.stringify(data)}</div>;
}

This is the canonical "fetch on user-id change" pattern. userId is closed over and listed in deps. r is a parameter of the .then() callback — local to that arrow function, not a closure.

Our rule fired:

React Hook useEffect has missing dependencies: r

Tracing into the source, extractLocallyDeclaredIdentifiers walked the effect body, collected VariableDeclaration and FunctionDeclaration names, but didn't collect params of nested ArrowFunctionExpression / FunctionExpression. Every callback parameter inside the effect was treated as a closure-from-outside.

Fix: when visiting a nested function node, add its params to the declared set:

if (
  n.type === "ArrowFunctionExpression" ||
  n.type === "FunctionExpression" ||
  n.type === "FunctionDeclaration"
) {
  for (const param of n.params) collectFromPattern(param);
}

collectFromPattern handles Identifier, ObjectPattern (with nested Property and RestElement), ArrayPattern, RestElement, and AssignmentPattern — destructured params, rest spreads, defaults. After the fix, the fixture passes.

The reason unit tests missed this: every test fixture in the suite used either a closure-only effect or an effect with a single top-level callback. None had .then((r) => …).then((data) => …) — the most common real-world shape.

Bug #2: NoSQL injection via $where was invisible

The fixture:

async function searchByName(req) {
  return db
    .collection("items")
    .find({
      $where: `this.name == '${req.query.name}'`,
    })
    .toArray();
}

This is a real NoSQL injection. $where evaluates JavaScript on the database server. With req.query.name interpolated unescaped, an attacker sends name=' || true || ' and gets every record.

Our rule didn't fire. Walking the source:

function getNodeSource(node: TSESTree.Node): string {
  if (node.type === Identifier) return node.name;
  if (node.type === MemberExpression) /* …recurse */ ;
  if (node.type === Literal) return String(node.value);
  return "[expression]"; // ← TemplateLiteral hit this
}

function containsUserInput(node: TSESTree.Node): boolean {
  const code = getNodeSource(node);
  return USER_INPUT_PATTERNS.some((pattern) => code.includes(pattern));
}

When the value of $where was a TemplateLiteral, getNodeSource returned the literal string '[expression]'. Then containsUserInput checked whether '[expression]' contained req.query — it doesn't. Silent skip.

The fix is to recurse into composite expressions instead of stringifying them:

function containsUserInput(node: TSESTree.Node): boolean {
  if (node.type === TemplateLiteral) {
    return node.expressions.some(containsUserInput);
  }
  if (node.type === BinaryExpression) {
    return containsUserInput(node.left) || containsUserInput(node.right);
  }
  if (node.type === CallExpression) {
    return (
      containsUserInput(node.callee) ||
      node.arguments.some(
        (a) => a.type !== "SpreadElement" && containsUserInput(a),
      )
    );
  }
  if (node.type === MemberExpression) {
    return USER_INPUT_PATTERNS.some((p) => getNodeSource(node).includes(p));
  }
  return false;
}

TemplateLiteral, BinaryExpression (string concat), and CallExpression (e.g. .toString() chains, String(req.x), JSON.stringify(req.body)) are all routes for tainted data into a query. Each gets recursed into now.

Why the unit tests missed it: the existing test corpus had find({ x: req.body.x }) shapes — direct user input as a property value. That shape gets caught by isUnsafePropertyValue's MemberExpression branch. The $where template literal is also user input, but expressed differently — and the pattern-matching code path didn't recurse far enough to see it.

Bug #3: AI-output detection missed the standard SDK pattern

The fixture:

import { generateText } from "ai";
import { openai } from "@ai-sdk/openai";

async function render(prompt: string, target: HTMLElement) {
  const { text } = await generateText({ model: openai("gpt-4"), prompt });
  target.innerHTML = text; // ← LLM output flows directly into innerHTML
}

This is straight from the Vercel AI SDK's official documentation. const { text } = await generateText(...) is the destructured pattern every example uses.

Our rule fired on nothing. The detection model:

const aiOutputPatterns = [
  "result.text",
  "response.text",
  "completion",
  "generated",
  "aiOutput",
  "aiResponse",
  "llmOutput",
  ".text",
];

function isLikelyAIOutput(node: TSESTree.Node): boolean {
  const text = sourceCode.getText(node);
  return aiOutputPatterns.some((pattern) => text.includes(pattern));
}

When the rule visits target.innerHTML = text, the right-hand side is the bare identifier text. The string 'text' doesn't match 'result.text', 'response.text', or '.text' (which all require a member-access prefix). So isLikelyAIOutput returns false. No diagnostic.

The pattern list assumes the LLM result is referenced as a property of an object. But the destructured pattern produces a free identifier. Two completely valid sources, only one detectable.

The fix is to add scope tracking — record any local variable bound from a known AI SDK call, and treat references to those as AI output:

const aiBoundNames = new Set<string>();
const AI_SDK_CALLS = new Set([
  "generateText",
  "streamText",
  "generateObject",
  "streamObject",
]);

function isAISDKCall(node: TSESTree.Expression): boolean {
  let target = node;
  if (target.type === "AwaitExpression") target = target.argument;
  if (target.type !== "CallExpression") return false;
  const callee = target.callee;
  if (callee.type === "Identifier" && AI_SDK_CALLS.has(callee.name))
    return true;
  if (
    callee.type === "MemberExpression" &&
    callee.property.type === "Identifier" &&
    AI_SDK_CALLS.has(callee.property.name)
  )
    return true;
  return false;
}

return {
  VariableDeclarator(node) {
    if (!node.init || !isAISDKCall(node.init)) return;
    if (node.id.type === "Identifier") {
      aiBoundNames.add(node.id.name);
    } else if (node.id.type === "ObjectPattern") {
      for (const prop of node.id.properties) {
        if (prop.type === "Property" && prop.value.type === "Identifier") {
          aiBoundNames.add(prop.value.name);
        }
      }
    }
  },
  // …
};

Now both const result = await generateText(...) (binding result → access via result.text) and const { text } = await generateText(...) (binding text directly) flow into aiBoundNames. The isLikelyAIOutput check picks them up by referenced identifier, regardless of how the user destructured.

Why the unit tests missed it: the test corpus used result.text patterns, matching 'result.text' in the patterns list literally. The destructured pattern was never in the test suite — even though it's the more common shape in production code.

The uncomfortable part: two of these three blind spots are exactly what AI writes

Re-read the fixtures with one question in mind: what does an AI assistant produce when you ask it for this?

• Ask any model for "fetch user data in a useEffect" and you get .then((r) => r.json()) — the inner-callback shape that bug #1 false-positived on.
• Ask for "render the model's response into the page" and you get const { text } = await generateText(...) straight from the Vercel AI SDK docs, then el.innerHTML = text — the destructure-into-sink that bug #3 was blind to.

That's not a coincidence. AI assistants are trained on the canonical documentation examples, so they emit the canonical shapes — the same shapes a rule author, writing tests from their own mental model, is least likely to enumerate. The result is a specific, repeatable failure: the rule has its widest blind spot exactly where AI-generated code is densest. A team that adopts a security linter to backstop AI output is, by default, trusting the rule on the one input distribution it was never tested against.

This is the same thread I keep pulling on from the other direction — I let Claude write 80 functions and 65–75% had security vulnerabilities, and Claude wrote a NestJS service; ESLint found 6 security holes. There the AI shipped the vulnerability. Here the AI ships a vulnerability the linter silently waves through. Both failure modes only surface when your fixtures are written from the patterns code actually takes — not the ones you imagined.

And the shape of the AI output is not uniform across models, which makes this worse, not better. When I ran 700 AI-generated functions through 5 models and ranked them by security, the rankings inverted the moment I broke them down by domain: the "most dangerous" model fixes 93% of the database vulnerabilities it writes, the "safest" fixes 45%. Different models favor different canonical shapes — so the same rule that's blind to Claude's preferred const { text } = await generateText(...) may sail through a different model's result.text and trip on a third's. A linter validated against one model's output distribution is not validated against the next model you swap in. The corpus is the only thing that holds the line, because its fixtures come from documentation — the source every model is trained on — not from whichever assistant you happened to test with.

If you want to reproduce this without our corpus: paste the three fixtures above into a file, point your model of choice — Claude, GPT, Gemini, whatever your team ships on — at the same three prompts, and diff what it generates against what your linter flags. Run it across two or three models and the blind spots move; the gap between "what the model writes" and "what your linter catches" is the test you were missing. (This is also the cleanest way to extend the benchmark above to a new model: same prompts, same rules, one column added.)

The two AI-shape detections that shipped with these patches — the destructured-generateText sink and the $where-template injection — are turned on by the single install + config block at the bottom of this article.

The one line worth pasting into your team channel: a security linter you adopted to backstop AI is, by default, untested on exactly the shapes that AI emits most.

Run it on Gemini: the blind spot moves with the model

"Point your model of choice" is not a hand-wave — the model you pick changes which of these three fixtures the linter ends up tested against, because each model favors different canonical shapes. I have the cross-model data to show it, from the same benchmark harness these rules ship with.

Take bug #2, the $where template injection — a database-layer NoSQL flaw. When I ran the database prompts across five models, Gemini 2.5 Pro topped the Database Operations domain at a 96% vulnerability rate on generation, yet fixes 93% of the database vulnerabilities it writes when asked — the highest remediation rate of any model tested. (Full per-domain table: Aggregate Benchmarks Lie.) Translation for bug #2: Gemini will hand you the $where-template shape readily, and a rule blind to template-literal recursion waves it straight through — the exact failure this corpus caught. The fixture is the model-agnostic backstop precisely because both Claude and Gemini are trained on the same MongoDB docs that demonstrate $where.

Now run the article's own recipe on Gemini and score what it generates against the three patched rules. Each row below pairs the documented Gemini 2.5 default shape for that prompt with the rule verdict that shape deterministically triggers, before and after the patch:

The shapes in column two are the documented Gemini-CLI defaults — the same family I observed when I ran the identical NestJS prompt on Claude and Gemini (Gemini 2.5 Flash, n=1 per toolchain), and the same database-shape tendency the Claude-vs-Gemini dead-heat run quantified across four domains. The verdict columns are deterministic — they are the rule's behavior on those shapes before and after the patch, the same F1 the smoke gate computes.

The honest caveat, because it is the whole point of the piece: I have the per-domain Gemini generation rates at n=700, but a fixture-by-fixture Gemini run against all three prompts at statistical n is its own experiment — the companion writeup, not this one. What is not in doubt is the direction: swap the model and the blind spots relocate, while the documentation-sourced fixture holds the line for every model because it is downstream of the docs they all trained on. That is the case for gating on the corpus instead of on whichever assistant you happened to test with.

What this whole episode is really about

Three rules. Three bugs. All caught by ground truth, none by unit tests. The pattern across them is the same:

Unit tests verify that the rule does what its author thought it should do. The author wrote the test, the author writes the rule, the same mental model produces both. If the author didn't think of the .then((r) => …) pattern, neither the rule nor the tests cover it. The tests pass; the rule has a hole.

Ground-truth corpora verify that the rule does what the world needs it to do. The fixtures are written from real CVE shapes, real framework documentation, real production codebases. They don't match the rule's mental model — they match the user's. Mismatches surface as F1<1.00.

And this is why these bugs survived code review, not just unit tests. When I reviewed each of these rules, I had two green signals in front of me: a passing test suite and a clean run against peer plugins on real OSS. Two independent checks, both green — that is normally enough to approve. What I couldn't see from the diff was that both signals shared the same blind spot. The unit tests encoded the author's mental model; the OSS sweep happened not to contain the $where-template or destructured-generateText shapes in the sampled files. Two green checks that fail the same way look exactly like two green checks that pass. The corpus was the first signal with an independent source of truth — documentation, not the author — so it was the first one that could disagree.

The fixtures in our suite are tiny — 12 to 18 lines per corpus, 4 fixtures each. The total disk cost is under 5KB. They run in ~3 seconds total. They caught three bugs the unit tests had missed across months of development.

A 5KB corpus that runs in 3 seconds found three bugs that green unit tests had carried for months. That should change how you think about "what does it mean to test a static-analysis rule."

Three concrete takeaways for any team writing or shipping linters:

Write fixtures from documentation, not from your tests. When you start a new rule, open the canonical docs for the pattern (CVE description, framework doc, OWASP example). Copy the example into a fixture before writing the rule. If the rule passes the fixture later, you've shipped a feature; if it doesn't, you've found a bug before users do.

Make the corpus a CI gate. Unit tests verify implementation; corpus tests verify behavior. Treating them as the same kind of test means one of them will atrophy. Run both, fail the build on either.

Surface the failures with confusion-matrix detail. "Test failed" is one bit. "F1 = 0.67, TP=1 FP=0 FN=1 TN=2 — where-string.js did not fire" is the actual diagnostic. The test framework should output the matrix, not just the boolean. Triage time goes from 15 minutes to 30 seconds.

The config

If you just want the three fixed rules in your own pipeline — including the $where and destructured-generateText detection that shipped with these patches — install the plugins and turn the rules on:

`bash
npm i -D eslint-plugin-react-features eslint-plugin-mongodb-security eslint-plugin-vercel-ai-security
`

`js
// eslint.config.js (flat config)
import reactFeatures from "eslint-plugin-react-features";
import mongodbSecurity from "eslint-plugin-mongodb-security";
import vercelAiSecurity from "eslint-plugin-vercel-ai-security";

export default [
{
plugins: {
"react-features": reactFeatures,
"mongodb-security": mongodbSecurity,
"vercel-ai-security": vercelAiSecurity,
},
rules: {
"react-features/hooks-exhaustive-deps": "error",
"mongodb-security/no-unsafe-query": "error",
"vercel-ai-security/no-unsafe-output-handling": "error",
},
},
];
`

New to these plugins? The per-rule docs and quick-starts live here: vercel-ai-security and mongodb-security.

The three fixes here are in packages/eslint-plugin-react-features, eslint-plugin-mongodb-security, and eslint-plugin-vercel-ai-security. The corpora are in benchmarks/corpus/. The smoke gate is benchmarks/suites/ilb-flagship/smoke.ts and it runs in three seconds.

Three seconds. Three bugs. Months of "fully tested." Pick which signal you trust.

What's the bug that got past two green signals at once on your team — the one where the tests and the second check were both green because they shared the same blind spot, and a single fixture copied from the docs would have caught it the day it shipped? I want to hear it in the comments.

Two more from the same bench, written up separately

This piece is part of the Inside our linter benchmarks series — the smoke gate caught the three above. The full ILB-Flagship sweep on 45K+-star OSS repos exposed the same class of bug from the opposite direction, in two more rules:

• 5 cycles invisible in 14,556 files — the cache bug that hid them (and the cache-poisoning-at-scale companion) — our import-next/no-cycle reported 0 cycles on next.js's 14K-file repo, but found 5+ in a 33-file subset of the same repo, same rule, same config. The bug was a DFS cache that encoded "depth limit hit" as "proven acyclic," and the cascade swallowed 245 files. A false negative, where these three were false negatives and a false positive.
• The false-positive twin: no-hardcoded-credentials fired 842 times on vercel/ai; 807 were TypeScript union-type literals and error class names, not secrets. Same root cause as bug #2 here — a detector matching on shape without enough context — just inverted. (Writeup pending; the benchmark methodology behind both is in I ran 40 vulnerable patterns through 17 plugins, which also covers what these rules do on AI-generated code.)

All of these survived months of unit-test coverage. All fell to ground-truth fixtures + bench data. Same lesson, more receipts: the test that disagrees with the author is the only one that can find the author's blind spot.

📊 About the author

I'm Ofri Peretz, building the Interlace ESLint ecosystem — a JavaScript static-analysis catalog that runs under ESLint and Oxlint with CI-enforced parity.

• 🔗 Portfolio & live metrics
• 📦 The Interlace ESLint plugins on npm
• 🐙 GitHub: ofri-peretz/eslint
• 📈 Live impact dashboard

Ofri PeretzFollow

IC5/M2 Leader @ Snappy US. Building revenue APIs & AI-ready ESLint plugins. Expert in distributed teams, scalable infra, and fostering a culture of craftsmanship.