AI-era security scanner for LLM-generated code.
Detect hallucinated vulnerabilities, auth bypasses, floating promises, and risky AI-generated patterns in TypeScript, JavaScript, Python, and Go — before production.
Aegis prioritizes precision over noisy detection. We aim for zero false positives.
# Cargo (recommended)
cargo install aegis
# npm
npm install -g aegis-security
aegis audit .No config needed. No cloud. No API keys. 100% local.
You can ignore specific directories using a .aegisignore file in the root of your project.
To ignore specific rules inline, add an aegis-ignore comment on the violation line or the line directly above it:
// aegis-ignore: ai-go-insecure-tls
tr := &http.Transport{
TLSClientConfig: &tls.Config{InsecureSkipVerify: true},
}Or ignore all rules for a specific block:
// aegis-ignore: all
eval(userInput);| Language | Rules | Status |
|---|---|---|
| TypeScript / TSX | 14 | Stable |
| JavaScript / JSX | 7 | Stable |
| Python | 12 | Stable |
| Go | 5 | Stable |
| Rust | — | Planned |
More languages coming soon. Aegis uses Tree-sitter, which supports 40+ languages — the engine is ready, we just need community-driven rules.
$ aegis audit .
[HIGH] Possible Auth Bypass. Sensitive route lacks authorization (role) checks. in auth.ts:35
Explanation: AI often implements JWT verification but forgets to check
if the user actually has the privileges to access the route.
[HIGH] Floating promise detected. AI often forgets to await async functions. in db.ts:12
Explanation: Unawaited database calls may continue executing after
the request lifecycle ends, causing race conditions.
[HIGH] Hardcoded secret detected. in config.ts:3
Explanation: AI coding assistants frequently insert mock API keys
which developers forget to move into environment variables.
[MEDIUM] Fake validation detected. in users.ts:7
Explanation: AI generates if-checks that log errors but never return,
allowing execution to continue past failed validation.
Scan Summary
────────────
HIGH: 8
MEDIUM: 1
LOW: 0
AI Risk Score: 82/100
[!] Found 9 violationsAI coding assistants (Copilot, Cursor, Claude, DeepSeek) have drastically increased development speed. But they introduced a new class of bugs that traditional linters completely miss — because the syntax is perfectly valid. The logic is what's broken.
Aegis doesn't ask "is this code vulnerable?" Aegis asks "did an AI write this, and did it cut corners?"
We use Tree-sitter for AST parsing and a custom YAML rule engine to catch these behavioral security flaws with surgical precision.
These are actual patterns that LLMs produce daily. Aegis catches all of them.
Forgotten await (Race Condition)
app.post('/api/users', (req, res) => {
db.save(req.body); // ← No await. Data may never persist.
res.send("Saved");
});Fake Validation (Auth Bypass)
if (!req.body.email) {
console.log("Email missing"); // ← No return. Code continues below.
}
createUser(req.body); // ← Runs even without emailSilent Error Swallowing
try {
await chargeCustomer(order);
} catch (e) {
console.log(e); // ← Payment fails silently.
}SQL Injection via f-string
# AI's #1 Python mistake — f-strings in SQL
cursor.execute(f"SELECT * FROM users WHERE id = {user_id}")Unsafe Pickle Deserialization (RCE)
# AI suggests pickle for everything — allows remote code execution
model = pickle.load(open("model.pkl", "rb"))DEBUG=True Left in Production
# AI generates Django/Flask boilerplate and never flips this
DEBUG = True
SECRET_KEY = "super-secret-key-12345"Unsafe eval()
// AI sometimes uses eval for dynamic execution
const userInput = req.body.code;
eval(userInput); // ← Remote code execution riskHardcoded Secret
// AI inserts placeholder credentials that make it to production
const STRIPE_KEY = "sk_live_1234567890";Aegis ships with 38 precision-first rules across TypeScript, JavaScript, Python, and Go — all tested with zero false positives.
| Rule ID | Severity | Problem Detected |
|---|---|---|
ai-auth-bypass |
HIGH | JWT validated but no role/permission check |
ai-floating-promise |
HIGH | Unawaited async DB/Network calls |
ai-silent-fail |
HIGH | catch(e) { console.log(e) } |
ai-hardcoded-secret |
HIGH | Placeholder API keys left in code |
ai-regex-injection |
HIGH | new RegExp(req.query.q) |
ai-open-redirect |
HIGH | res.redirect(req.query.next) |
ai-insecure-fetch |
HIGH | rejectUnauthorized: false |
ai-missing-rate-limit |
HIGH | Login/OTP route without brute-force protection |
ai-unsafe-innerhtml |
HIGH | innerHTML = userInput without sanitization |
ai-fake-validation |
MEDIUM | Validation block that never returns/throws |
ai-unsafe-deserialization |
HIGH | JSON.parse() without error handling |
ai-cors-wildcard |
HIGH | CORS configured with wildcard * origin |
ai-sql-injection-js |
HIGH | SQL query built via string concatenation |
ai-no-timeout |
HIGH | HTTP request without timeout configured |
ai-path-traversal |
HIGH | File ops with unsanitized user input |
ai-unhandled-rejection |
MEDIUM | Promise chain without .catch() handler |
| Rule ID | Severity | Problem Detected |
|---|---|---|
ai-hardcoded-secret-js |
HIGH | Credentials hardcoded in JS source |
ai-unsafe-eval |
HIGH | eval() or new Function() usage |
ai-no-csrf |
HIGH | POST/PUT/DELETE routes without CSRF protection |
ai-weak-crypto |
HIGH | MD5, SHA1, or DES algorithms detected |
ai-missing-auth-check |
HIGH | Route handler without auth middleware |
ai-nosql-injection |
HIGH | MongoDB query with raw user input |
ai-insecure-jwt |
HIGH | JWT with weak/none algorithm |
ai-exposed-stack-trace |
MEDIUM | Error handler exposes internal details |
ai-dangerously-set-html |
HIGH | React dangerouslySetInnerHTML without sanitization |
ai-missing-input-validation |
MEDIUM | Express route without body validation |
| Rule ID | Severity | Problem Detected |
|---|---|---|
ai-bare-except |
HIGH | except: that swallows all errors silently |
ai-hardcoded-secret-py |
HIGH | Credentials hardcoded in source files |
ai-sql-format-string |
HIGH | SQL injection via f-strings or .format() |
ai-debug-true |
HIGH | DEBUG = True left in Django/Flask settings |
ai-pickle-load |
HIGH | pickle.load() allows Remote Code Execution |
ai-hallucinated-import |
HIGH | AI-generated imports for non-existent packages |
ai-unsafe-yaml |
HIGH | yaml.load() instead of yaml.safe_load() |
ai-subprocess-shell |
HIGH | subprocess call with shell=True |
ai-plaintext-password |
HIGH | Password compared/stored in plaintext |
ai-assert-security |
HIGH | assert used for auth/security checks |
ai-requests-verify-false |
HIGH | TLS verification disabled in requests |
ai-default-secret-key |
HIGH | Django SECRET_KEY with placeholder value |
ai-broad-except |
MEDIUM | Catching Exception too broadly |
ai-django-debug-true |
HIGH | Django DEBUG=True in production |
| Rule ID | Severity | Problem Detected |
|---|---|---|
ai-go-sql-injection |
HIGH | SQL query built via string formatting or concatenation |
ai-go-insecure-tls |
HIGH | TLS verification disabled via InsecureSkipVerify |
ai-go-swallowed-error |
MEDIUM | Error swallowed via blank identifier _ |
ai-go-command-injection |
HIGH | OS command arguments built using concatenation |
ai-go-hardcoded-secret |
HIGH | API keys and credentials hardcoded in Go source |
Rust and Tree-sitter make Aegis blazingly fast.
Scanned 12,000 LOC in 0.08s
Aegis adds zero noticeable overhead to your CI/CD pipeline.
Run the benchmark suite:
cargo benchAegis includes a comprehensive test suite covering the engine, parser, and rule modules.
cargo testTest coverage includes:
- Language detection (TypeScript, JavaScript, Python, unknown)
- AST parsing (valid and invalid syntax)
- Rule loading (YAML files, directory traversal, error handling)
- SARIF output generation
- Confidence level serialization
- File scanning with real tree-sitter queries
"Isn't this just Semgrep?" — No. We share architectural inspiration (AST + YAML rules), but we solve a fundamentally different problem.
| Semgrep | Aegis | |
|---|---|---|
| Focus | Generic SAST | AI behavioral security |
| Audience | Security Engineers | Developers using AI Assistants |
| Detects | General vulnerabilities (SQLi, XSS) | LLM-specific failure patterns |
| Strategy | Broad detection | Precision-first (Zero false-positive goal) |
| Output | Standard warnings | Explanation Engine (Why the AI made this mistake) |
Drop Aegis into your CI/CD pipeline. It automatically audits every Pull Request and uploads SARIF results to GitHub's Security tab.
name: Aegis Security Audit
on: [push, pull_request]
jobs:
audit:
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@v4
- uses: vahapogut/aegis-action@v1
with:
format: 'sarif' # Results appear in Security → Code ScanningOr use it directly in any CI:
# Exit code 1 if HIGH violations found = blocks merge
aegis audit . --format sarif > aegis-results.sarifAegis supports JSON and SARIF v2.1.0 output for easy integration with GitHub Security tab, Datadog, or custom dashboards.
aegis audit . --format text # default, colored terminal output
aegis audit . --format json # machine-readable with risk score
aegis audit . --format sarif # GitHub Security tab integrationAegis rules are YAML files with Tree-sitter queries. See CONTRIBUTING.md for the full guide.
rule:
id: "ai-your-pattern"
language: "typescript"
confidence: "HIGH"
message: "Short description of the bug."
explanation: "Why the AI does this and how to fix it."
query: >
(catch_clause
body: (statement_block) @block
(#not-match? @block "throw|return")
)Apache-2.0. Built for the community by IPEC Labs.