/sentinel

Vulnerability Detection Patterns

Purpose: Use this reference during SCAN to identify common vulnerability patterns, confirm likely findings, and compare safe versus unsafe code.

Interpretation Rules

• Regexes are heuristics, not proof.
• Confirm source-to-sink reachability before escalating a finding.
• For generic secrets, combine regex, entropy, and contextual review.
• AI-generated code is 2.74x more likely to contain XSS than human code — apply extra scrutiny.

Hardcoded Secrets

Detection Patterns

# API Keys
(api[_-]?key|apikey)['":\s]*[=:]\s*['"][a-zA-Z0-9]{20,}['"]

# AWS Keys
(AKIA|ABIA|ACCA|ASIA)[A-Z0-9]{16}

# Private Keys
-----BEGIN (RSA|DSA|EC|OPENSSH) PRIVATE KEY-----

# Generic Secrets (entropy + context review recommended)
(password|passwd|pwd|secret|token)['":\s]*[=:]\s*['"][^'"]{8,}['"]

# GitHub Tokens
(ghp|gho|ghu|ghs|ghr)_[A-Za-z0-9_]{36,}

# JWT tokens
eyJ[A-Za-z0-9_-]{10,}\.eyJ[A-Za-z0-9_-]{10,}\.[A-Za-z0-9_-]+

Bad: const API_KEY = 'sk_live_abc123xyz789'; Good: const API_KEY = process.env.API_KEY;

Regex limitation: known formats (AWS, GitHub tokens) are reliable; generic secrets need entropy checks and contextual review.

SQL Injection

(query|execute|exec)\s*\(\s*['"`].*\$\{.*\}
(query|execute|exec)\s*\(\s*['"`].*\+\s*\w+
SELECT.*FROM.*WHERE.*['"`]\s*\+
INSERT.*INTO.*VALUES.*['"`]\s*\+

Bad: db.query(\SELECT * FROM users WHERE id = ${userId}`);**Good:**db.query('SELECT * FROM users WHERE id = ?', [userId]);`

XSS (Cross-Site Scripting)

\.innerHTML\s*=\s*[^'"]
dangerouslySetInnerHTML
document\.write\s*\(
\.html\s*\(\s*[^'"]

Bad: element.innerHTML = userInput; Good: element.textContent = userInput; or element.innerHTML = DOMPurify.sanitize(userInput);

AI-generated code fails XSS tests 86% of the time — flag all innerHTML assignments in AI-assisted code.

Command Injection

exec\s*\(\s*['"`].*\$\{
spawn\s*\(\s*['"`].*\+
child_process.*\$\{
os\.system\s*\(.*\+|os\.system\s*\(.*f['"]
subprocess\.(call|run|Popen)\s*\(.*shell\s*=\s*True

Bad: exec(\ls ${userDir}`);**Good:**execFile('ls', [sanitizedDir]);`

Path Traversal

(readFile|writeFile|unlink)\s*\(.*\+.*\)
path\.join\s*\(.*req\.(params|query|body)
\.\.\/

Bad: fs.readFile(baseDir + req.params.file); Good:

const safePath = path.join(baseDir, path.basename(req.params.file));
if (!safePath.startsWith(baseDir)) throw new Error('Invalid path');
fs.readFile(safePath);

NoSQL Injection

\$where|\$gt|\$lt|\$ne|\$regex|\$exists
(find|findOne|aggregate|update)\s*\(\s*\{.*req\.(body|query|params)

Bad: db.collection('users').find({ username: req.body.username, password: req.body.password }); Good: app.use(mongoSanitize());

Prompt Injection

# LLM instruction override
(ignore|disregard|forget)\s+(previous|above|all)\s+(instructions|prompts|rules)
(system|admin)\s*:\s*(override|execute|run)
# Template injection in prompt construction
\$\{.*user.*\}.*prompt|prompt.*\$\{.*user
# Indirect injection markers
<\|im_start\|>|<\|im_end\|>|\[INST\]|\[/INST\]

Bad: const prompt = \Summarize this text: ${userInput}`;` Good:

const response = await llm.complete({
  system: 'You are a text summarizer. Only summarize the provided text.',
  user: userInput,
});

Prototype Pollution

__proto__|constructor\s*\.\s*prototype|Object\.assign\s*\(\s*\{\}
(merge|extend|assign|defaults)\s*\(.*req\.(body|query|params)

Bad: for (const key in source) { target[key] = source[key]; } Good:

function safeMerge(target, source) {
  for (const key of Object.keys(source)) {
    if (key === '__proto__' || key === 'constructor' || key === 'prototype') continue;
    target[key] = source[key];
  }
}

Exceptional Conditions (OWASP A10:2025)

# Empty catch blocks
catch\s*\([^)]*\)\s*\{\s*\}
# Sensitive data in error messages (CWE-209)
(res\.status|res\.json|throw)\s*.*\.(stack|message|sql|query)
# Unhandled promise rejections
\.then\s*\([^)]+\)\s*;(?!\s*\.catch)
# Fail-open patterns (CWE-636)
catch\s*\([^)]*\)\s*\{[^}]*(continue|return\s+true|next\(\))

Bad: catch (e) { return res.status(500).json({ error: e.message, stack: e.stack }); } Good: catch (e) { logger.error('Operation failed', e); return res.status(500).json({ error: 'An error occurred' }); }

AI-Generated Code Signals

Flag these patterns with boosted severity when code is known or suspected to be AI-generated:

Signal Keywords → Recipe (SKILL.md excerpt)

For natural-language input without an explicit subcommand. Subcommand match wins if both apply.