← Back to the tool
Trace.
/

OPEN SOURCE

DETECTION ENGINE

Trace.

AI can write.

Trace can [read].

The first open-source security checker built for LLM failure modes.

Hallucinations. Credentials. Silent errors. Unsafe sanitization.

Try in browser→View on GitHub→
02 / 11LIVE SCANNER

How Trace reads the code your AI wrote.

A live scan of a 47-line FastAPI file. Detection engine runs on all 24 open-source patterns across 6 languages. Each issue is flagged with its severity, line, and a one-line explanation.

file.py · 47 lines · pythonCOMPLETE
01import fastapi
02from fastapi import FastAPI, Depends, HTTPException
03import fake_validator_xyz
04import sqlite3
05import jwt
06from pydantic import BaseModel
07
08app = FastAPI()
09
10# Config
11SECRET_KEY = "dev-jwt-secret-2026"
12STRIPE_KEY = "sk_live_51xZy_abcdef..."
13
14class User(BaseModel):
15 id: int
16 name: str
17
18def fetch_user(user_id):
19 conn = sqlite3.connect("db.sqlite")
20 q = f"SELECT * FROM users WHERE id = {user_id}"
21 return conn.execute(q).fetchone()
22
23@app.get("/users/{uid}")
24def get_user(uid: int):
25 try:
26 u = fetch_user(uid)
27 except Exception:
28 pass
29 return u
30
31@app.post("/users")
32def create_user(name: str):
33 conn = sqlite3.connect("db.sqlite")
34 conn.execute(
35 f"INSERT INTO users (name) VALUES ('{name}')"
36 )
37 conn.commit()
38
39@app.delete("/users/{uid}")
40def delete_user(uid: str):
41 eval(f"removeUser({uid})")
42
43@app.get("/search")
44def search(q: str):
45 return {"results": []}
46
47# end of file
5 detections found
✗criticalline 03hallucinated dependency
fake_validator_xyz · not on PyPI
✗criticalline 11credential leak
JWT secret hardcoded
✗criticalline 12credential leak
Stripe live key · redacted
✗criticalline 20SQL injection
f-string interpolation in execute
✗criticalline 41code injection
eval() with user input
03 / 11THE SHIFT

Nearly half of production code is now written by AI.

Yet the tools we use to secure it were built for human bugs — humans who hesitate, who review, who leave typos. AI writes confidently. That confidence is the attack surface.

0%

of AI-generated code contains security flaws.Veracode, 2025 State of Software Security

0%

of organizations found vulnerabilities in AI-generated code.CSA

0%

of committed code in 2025 was AI-written.SonarSource

04 / 11THE GAP

The failure mode is already loud.

Three incidents from the last three months. Nothing classified, nothing unusual. Ordinary AI-assisted deploys. Ordinary catastrophes.

MOLTBOOK · FEB 2026

1.5 million API keys exposed.

The founder wrote zero lines himself. AI scaffolded a Supabase database with public read/write permissions, and nobody reviewed the generated configuration before deploy.

AXIOS · MAR 2026

Supply chain attack via a maintainer token.

The compromised package was downloaded 450,000 times before detection. Downstream applications were affected within hours.

PYPI · APR 2026

Slopsquatting reaches 340 known packages.

Typo-squatted libraries matching the names that large language models hallucinate. Installation is one autocomplete away from code execution.

05 / 11WHY YOU

These are the quiet endings.

Not famous incidents. Not YC companies. Three things that happen to solo developers shipping AI-generated code without a second reader.

SCENARIO 01
02:00 AMSaturday night. Claude Code finished an auth flow that finally worked. You deployed it and went to sleep.
09:00 AMMonday morning in Japan. Someone hits /user?id='; DROP TABLE users;--
09:00 AMThe code Claude wrote was cursor.execute(f"SELECT * FROM users WHERE id = {user_id}"). The query ran.
11:00 AMEvery user in your database is gone. You hadn't set up automated backups yet.

The product you spent six weekends building no longer has users.

— Trace would have flagged the SQL injection before commit.

SCENARIO 02
DAY 1You asked Lovable to build a payment page. It wrote const stripe = new Stripe("sk_live_...") directly into the file.
DAY 1You pushed the repo to GitHub public, because that's the default.
DAY 1 +00:30Bots scanning GitHub for exposed Stripe keys found it in 30 minutes.
DAY 2Stripe notifies you of $47,000 in fraudulent charges. They refund everything.
DAY 3Stripe also closes your account for breach of security terms. You rebuild. You explain. It takes weeks.

Stripe refunded the money. Your reputation as a serious developer is already priced in.

— Trace would have refused the commit containing the live key.

SCENARIO 03
MONTH 01Claude suggested using react-form-validator-x to handle your signup form. You ran npm install. It failed. You tried again. It worked.
MONTH 01You didn't realize: someone had registered that exact name on npm 20 minutes before you tried the second time.
MONTH 04Your SaaS has 1,200 paying users. Revenue is growing. You're not watching dependencies anymore.
MONTH 05The package author pushes a new version. It exfiltrates your users' passwords on first run.
MONTH 05You find out from a Hacker News thread about your own product.

Every user password compromised. Trust reset to zero.

— Trace would have refused the import on day one. The package didn't exist in npm's registry yet.

06 / 11WHAT WE BUILT

Twenty-four patterns across six languages.

v0.7.0 ships 24 detection patterns across Python, JavaScript, TypeScript, Go, Rust, and Ruby. All open source. We catch the failure modes specific to AI-written code — not the ones humans are already careful about.

PYTHON / JAVASCRIPT / TYPESCRIPT
01HALLUCINATED DEPS● OSS

imports of packages that don't exist

02DEPRECATED APIS● OSS

calls to methods that don't exist

03CREDENTIAL LEAKS● OSS

api keys, tokens, connection strings

04FAKE TYPE SAFETY● OSS

any abuse, stripped generics

05SILENT EXCEPTIONS● OSS

empty catches, swallowed promises

06UNSAFE SANITIZE● OSS

SQL injection, XSS, command inject

07TAUTOLOGICAL TEST● OSS

tests that can never fail

GO
08SLOPSQUATTING● OSS

suspicious import paths that may be AI-hallucinated

09ERROR IGNORED● OSS

error return values explicitly discarded with _

10SPRINTF SQL● OSS

SQL queries built with fmt.Sprintf or string concat

11HARDCODED SECRET● OSS

API keys, tokens, credentials in Go source

RUST
12UNWRAP ABUSE● OSS

excessive .unwrap() that can panic at runtime

13UNSAFE BLOCK● OSS

unsafe blocks bypassing safety guarantees

14TODO MACRO● OSS

todo!()/unimplemented!() placeholders that panic

15PANIC MACRO● OSS

panic!() where Result<T, E> should be used

RUBY
16MASS ASSIGNMENT● OSS

ActiveRecord mass assignment without strong params

17SQL INTERPOLATION● OSS

string interpolation inside SQL queries

18SILENT RESCUE● OSS

rescue blocks that silently swallow exceptions

19EVAL INJECTION● OSS

eval/send called with dynamic input

CROSS-LANGUAGE · NEW IN v0.7.0
20MISSING AWAIT● OSS

async call assigned without await — value is a Promise, not the result

21INSECURE RNG● OSS

Math.random() or random.random() used for security-sensitive values

22DYNAMIC EVAL● OSS

eval(), new Function(), or exec() with variable argument — code injection

23HARDCODED LOCALHOST● OSS

http://localhost or http://127.0.0.1 in non-test code

24ENV WITHOUT FALLBACK● OSS

os.environ["X"] or process.env.X without fallback — crashes on missing var

Open the full engine on GitHub →
07 / 11AI DEMO

Watch AI write code. Then watch Trace read it.

Pick a prompt. See real AI-generated code. Trace analyzes it on our servers — catching what humans miss.

08 / 11USAGE

How to use it.

Three ways to run Trace, sorted by effort. Pick the one that fits where you write code.

METHOD 01 · BROWSER

No install. No signup.

Open tracecheck.dev.
Paste your code. Press Check.
Results in about 400 milliseconds.

Open Playground →

METHOD 02 · COMMAND LINE

Requires Node.js.

$ npx trace-core your-file.py

npx ships with Node.js. If you don't have Node.js yet, download it at nodejs.org.

METHOD 03 · CONTINUOUS INTEGRATION

Blocks commits with issues.

# .github/workflows/check.yml
name: Trace
on: [push, pull_request]
jobs:
  trace:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - run: npx trace-core src/

Drop into any Git provider's CI. Detection failures cause the workflow to exit with code 1.

09 / 11LANGUAGES

Six languages. Twenty-four patterns.

Full detector coverage with language-specific patterns. More languages shipping soon.

Python
JavaScript
TypeScript
Go
Rust
Ruby

Coming soon: Java · PHP · Dart

10 / 11START

Three ways in.

BROWSER

For quick checks. No install.

CLI

For CI and pre-commit. npm.

$ npx trace-core your-file.py
GITHUB

For the full source. MIT licensed.