opengrep-rule-generator

Opengrep Rule Generator

Overview

Generate valid opengrep/semgrep YAML rules through collaborative dialogue. Supports two workflows: guided (interactive Q&A to discover what to detect) and vulnerability-driven (given CVEs, OWASP categories, or vulnerability descriptions, generate rules automatically).

When to Use

• User says "create a rule", "write a rule", "generate a rule", "detect [vulnerability]"
• User provides a CVE, CWE, or OWASP reference and wants detection rules
• User shares code snippets and asks "how do I catch this pattern?"
• User wants to scan a codebase for a class of vulnerabilities
• User asks to audit code for security issues and wants reusable rules

Process Flow

digraph rule_gen {
    "User request" [shape=doublecircle];
    "Has specific vulnerability?" [shape=diamond];
    "Guided Discovery" [shape=box];
    "Vulnerability-Driven" [shape=box];
    "Gather context" [shape=box];
    "Choose rule mode" [shape=diamond];
    "Generate Search rule" [shape=box];
    "Generate Taint rule" [shape=box];
    "Present rule for review" [shape=box];
    "User approves?" [shape=diamond];
    "Write rule file" [shape=box];
    "Generate test file" [shape=box];
    "Validate with opengrep" [shape=doublecircle];

    "User request" -> "Has specific vulnerability?";
    "Has specific vulnerability?" -> "Vulnerability-Driven" [label="yes"];
    "Has specific vulnerability?" -> "Guided Discovery" [label="no"];
    "Guided Discovery" -> "Gather context";
    "Vulnerability-Driven" -> "Gather context";
    "Gather context" -> "Choose rule mode";
    "Choose rule mode" -> "Generate Search rule" [label="pattern match"];
    "Choose rule mode" -> "Generate Taint rule" [label="data flow"];
    "Generate Search rule" -> "Present rule for review";
    "Generate Taint rule" -> "Present rule for review";
    "Present rule for review" -> "User approves?";
    "User approves?" -> "Present rule for review" [label="revise"];
    "User approves?" -> "Write rule file" [label="yes"];
    "Write rule file" -> "Generate test file";
    "Generate test file" -> "Validate with opengrep";
}

Reference Documentation

Before generating any rule, load context from these project docs:

Doc	Path	Load When
Rule syntax & templates	docs/ai/RULES_SYNTAX.md	Always — contains all operators, templates, language IDs
Rule index	docs/ai/RULES_INDEX.md	When checking for existing similar rules
Engine internals	docs/ai/RULES_ENGINE.md	When choosing between search vs taint mode

Also search existing rules for similar patterns:

• semgrep-rules/semgrep-rules/<language>/ — community rules
• semgrep-rules-trailbits/<language>/ — Trail of Bits rules

Workflow 1: Guided Discovery

Ask these questions one at a time to scope the rule:

Step 1 — What are we detecting?

What vulnerability, bug pattern, or coding anti-pattern do you want to detect?

Examples:
- "SQL injection in our Flask app"
- "Hardcoded secrets in config files"
- "Missing null checks after API calls"
- "Insecure deserialization"
- "Race conditions in Go goroutines"

Step 2 — What language and frameworks?

What programming language(s) and frameworks are involved?

I support 30+ languages including: python, javascript, typescript, java, go,
ruby, php, csharp, c, rust, scala, kotlin, swift, terraform/hcl, yaml,
dockerfile, solidity, and more. I also support 'generic' for config files
and 'regex' for raw text matching.

Step 3 — Show me the vulnerable code

Can you show me an example of the VULNERABLE code you want to catch?
And if possible, also show the SAFE version (the fix)?

This helps me write precise patterns with fewer false positives.

Step 4 — Where does user input enter? (if applicable)

Does this vulnerability involve untrusted user input flowing into a dangerous function?

If yes, I'll use TAINT MODE (tracks data flow from sources to sinks).
If no, I'll use SEARCH MODE (structural pattern matching).

For taint mode, I need to know:
- Sources: Where does untrusted data come from? (HTTP requests, file reads, env vars, etc.)
- Sinks: Where is it dangerous? (SQL queries, system commands, file writes, etc.)
- Sanitizers: What makes the data safe? (escaping, validation, type casting, etc.)

Step 5 — Severity and scope

How severe is this issue?
- CRITICAL/ERROR: Exploitable vulnerability, must fix
- WARNING/MEDIUM: Likely vulnerability, should fix
- INFO/LOW: Code smell or audit flag

Should the rule apply to all files, or specific paths only?

Workflow 2: Vulnerability-Driven Generation

When given CVEs, CWEs, OWASP categories, or vulnerability descriptions:

1. Research the vulnerability — use the research plan below to gather full context
2. Parse the vulnerability — extract: affected language, vulnerable API/pattern, attack vector, fix
3. Check existing rules — search semgrep-rules/ and semgrep-rules-trailbits/ for coverage
4. Determine rule mode:
   • Data flows from input to dangerous function → Taint mode
   • Dangerous function call or config pattern → Search mode
   • Text/config pattern without AST → Regex/generic mode
5. Generate rule(s) — may produce multiple rules for different attack variants
6. Present with explanation of what each rule catches and known limitations

Vulnerability Research Plan

Before generating rules, conduct research to ensure comprehensive coverage. Use WebSearch and WebFetch tools.

Phase 1 — Understand the vulnerability:

• Search "<CVE/CWE ID>" vulnerability details to get official descriptions
• Search "<vulnerability type>" <language> exploit examples for real-world attack patterns
• Fetch the CWE entry from https://cwe.mitre.org/data/definitions/<ID>.html for taxonomy, related weaknesses, and detection methods
• Fetch the OWASP page for the relevant category (e.g., https://owasp.org/Top10/A03_2021-Injection/)

Phase 2 — Map language-specific attack surface:

• Search "<vulnerability>" <framework> cheat sheet site:cheatsheetseries.owasp.org
• Search "<vulnerability>" <framework> security best practices
• Identify: entry points (sources), dangerous APIs (sinks), safe alternatives (sanitizers)
• Search "<vulnerable function>" CVE to find real CVEs demonstrating the pattern

Phase 3 — Study existing detection:

• Search semgrep rule "<vulnerability type>" <language> for community rules
• Search existing rules in semgrep-rules/ and semgrep-rules-trailbits/ directories
• Note what's covered and what gaps remain

Phase 4 — Document findings: Write a brief research summary as a YAML comment block at the top of the rule file:

# Research: <vulnerability type> in <language/framework>
# Sources: <where untrusted data enters>
# Sinks: <where data becomes dangerous>
# Sanitizers: <what makes data safe>
# References researched:
#   - <url1>
#   - <url2>
# Coverage gaps found: <what existing rules miss>

Mode Selection Guide

Vulnerability Type	Rule Mode	Why
SQL injection	Taint	User input → query function
XSS	Taint	User input → HTML output
Command injection	Taint	User input → exec/system
SSRF	Taint	User input → HTTP request
Path traversal	Taint	User input → file operation
Hardcoded secrets	Search + regex	Pattern match on literals
Insecure config	Search	Structural pattern on config
Missing auth checks	Search (inside/not-inside)	Absence of pattern
Race conditions	Search (inside + not-inside)	Pattern within goroutine/thread
Weak crypto	Search	Specific API calls
Deserialization	Search or Taint	Depends on if input-controlled
Terraform misconfig	Search	HCL structural patterns

Rule Generation Template

When generating a rule, always produce this complete structure:

rules:
  - id: <company-or-project>-<language>-<vuln-type>
    message: >-
      <1-3 sentences: what was found, why it's dangerous, how to fix it>
    severity: <ERROR|WARNING|INFO>
    languages: [<language>]
    metadata:
      category: <security|correctness|best-practice|performance>
      cwe:
        - "CWE-XXX: <Description>"
      owasp:                          # include if security rule
        - "A0X:2021 - <Category>"
      references:
        - <url-to-documentation>
      technology:
        - <framework-or-library>
      subcategory:
        - <vuln|audit>
      confidence: <HIGH|MEDIUM|LOW>
      likelihood: <HIGH|MEDIUM|LOW>
      impact: <HIGH|MEDIUM|LOW>
    # ... pattern operators or taint spec ...

Naming Convention

Rule IDs: <scope>-<language>-<vuln-type>[-<variant>]

• myapp-python-sql-injection
• myapp-go-race-condition-map-write
• myapp-java-spring-ssrf

Message Quality

Good messages include:

1. What was detected (the pattern)
2. Why it's dangerous (the risk)
3. How to fix it (the remediation)

message: >-
  User input from `flask.request` is concatenated into a SQL query string
  passed to `cursor.execute()`. This could allow SQL injection, letting an
  attacker read or modify database contents. Use parameterized queries instead:
  `cursor.execute("SELECT * FROM users WHERE id = %s", (user_id,))`.

Test File Generation

Always generate a companion test file with:

# === True Positives (MUST trigger) ===

# ruleid: <rule-id>
<vulnerable code example 1>

# ruleid: <rule-id>
<vulnerable code example 2 — different variant>

# === True Negatives (must NOT trigger) ===

# ok: <rule-id>
<safe code — using the recommended fix>

# ok: <rule-id>
<safe code — different safe pattern>

Minimum: 2 true positives + 2 true negatives per rule.

Advanced Patterns Cookbook

Detect missing security check (pattern-not-inside)

patterns:
  - pattern: dangerous_operation(...)
  - pattern-not-inside: |
      if <... auth_check(...) ...>:
        ...

Detect tainted string concatenation (taint + metavariable-regex)

mode: taint
pattern-sources:
  - pattern: request.args.get(...)
pattern-sinks:
  - patterns:
      - pattern: |
          "$SQLSTR" + ...
      - metavariable-regex:
          metavariable: $SQLSTR
          regex: \s*(?i)(select|delete|insert|create|update|alter|drop)\b.*

Detect weak config values (metavariable-comparison)

patterns:
  - pattern-inside: |
      resource "aws_s3_bucket" "..." {
        ...
        versioning { days = $DAYS }
        ...
      }
  - metavariable-comparison:
      metavariable: $DAYS
      comparison: $DAYS < 90

Detect across multiple call variants (pattern-either + metavariable-regex)

patterns:
  - pattern-either:
      - pattern: $OBJ.$METHOD(...)
      - pattern: $MODULE.$METHOD(...)
  - metavariable-regex:
      metavariable: $METHOD
      regex: (eval|exec|compile|unsafe_load)

Detect with type constraints (typed metavariables)

# Go: match only http.Request types
pattern: ($REQ : *http.Request).$FIELD

Multi-step taint with labels

mode: taint
pattern-sources:
  - patterns:
      - pattern: request.get_param(...)
    label: USER_INPUT
  - patterns:
      - pattern: $X + $Y
    label: CONCATENATED
    requires: USER_INPUT
pattern-sinks:
  - patterns:
      - pattern: db.execute(...)
    requires: CONCATENATED

Reducing False Positives

After generating a rule, always consider adding:

1. pattern-not — exclude safe variants (pattern-not: func("...", ...) for literal strings)
2. pattern-not-inside — exclude safe contexts (inside try/catch, inside sanitization wrapper)
3. metavariable-regex — restrict metavariable values to dangerous ones
4. Safe type exclusions — options: { taint_assume_safe_numbers: true, taint_assume_safe_booleans: true }
5. Log/print exclusions — pattern-not-inside: $LOG.info(...) to avoid flagging logging

Validation Checklist

Before finalizing any rule:

• Rule ID is unique and follows kebab-case
• languages field matches the target code
• Message explains what, why, and how-to-fix
• Metadata includes CWE and OWASP where applicable
• At least 2 true positive test cases
• At least 2 true negative test cases
• False positive reducers considered (pattern-not, safe type options)
• Rule tested with opengrep scan --config <rule.yaml> <test-file> if available

Batch Generation

When asked to generate rules for a class of vulnerabilities (e.g., "OWASP Top 10 for Python Flask"):

1. List all applicable vulnerability categories
2. Check existing coverage in semgrep-rules/
3. Identify gaps
4. Generate rules for gaps only
5. Present as a table: | Vulnerability | Existing Rule? | New Rule ID | Mode |

Rule Generation Methodology

Follow this systematic process to generate high-quality rules:

1. Understand the Vulnerability Deeply

Don't just pattern-match on function names. Understand:

• What makes it dangerous? (the root cause, not the symptom)
• What's the attack scenario? (how does an attacker exploit this?)
• What's the data flow? (source → transformation → sink)
• What are ALL the variants? (string concat, format strings, template literals, f-strings)
• What are the safe alternatives? (parameterized queries, template engines, escaping functions)

2. Study Existing Rules as Templates

Before writing from scratch, search existing rules for similar patterns:

semgrep-rules/semgrep-rules/<language>/<framework>/security/
semgrep-rules-trailbits/<language>/

Read the best-quality rules and adapt their patterns. Existing rules show:

• Which sources are standard for each framework
• Which metavariable-regex patterns reduce false positives
• How pattern-not exclusions are structured
• What metadata fields and CWE/OWASP mappings to use

3. Build Source-Sink-Sanitizer Maps

For each language/framework, document:

Component	Examples
Sources (user input entry points)	request.args, req.query, $_GET, params[:]
Sinks (dangerous output points)	res.send(), echo, innerHTML, cursor.execute()
Sanitizers (safe transformations)	escape(), htmlspecialchars(), parseInt(), template engines
Propagators (data structure flow)	StringBuilder.append(), HashMap.put(), array.push()

4. Apply Defense-in-Depth Pattern Writing

Write rules in layers:

1. Broad taint rule — catches the primary attack vector
2. Specific search rules — catches dangerous API misuse (e.g., dangerouslySetInnerHTML)
3. Config/audit rules — flags risky configurations (e.g., autoescape off)

5. Iterate on False Positive Reduction

After initial rule generation:

1. Think through common safe patterns that would trigger
2. Add pattern-not for safe literals, safe types, logging
3. Add pattern-not-inside for try/catch, validation wrappers
4. Set taint_assume_safe_numbers: true and taint_assume_safe_booleans: true
5. Use metavariable-regex to restrict to dangerous method names

Full Context: Reference Documentation

When generating rules, these docs contain the complete rule schema:

What You Need	Where to Find It
All pattern operators	docs/ai/RULES_SYNTAX.md §3-5
Taint mode spec (sources/sinks/sanitizers/propagators/labels)	docs/ai/RULES_SYNTAX.md §6
Metavariable conditions (regex, comparison, type, pattern)	docs/ai/RULES_SYNTAX.md §4
All supported languages	docs/ai/RULES_SYNTAX.md §9
Rule templates (8 complete examples)	docs/ai/RULES_SYNTAX.md §12
How the engine matches patterns	docs/ai/RULES_ENGINE.md §2
How taint analysis works internally	docs/ai/RULES_ENGINE.md §3
Existing rule coverage by language	docs/ai/RULES_INDEX.md
CWE/OWASP metadata conventions	docs/ai/RULES_SYNTAX.md §7
Rule options (engine config)	docs/ai/RULES_SYNTAX.md §8

Generated Rules Directory

Store generated rules in custom-rules/<vuln-type>/ with companion test files:

custom-rules/
  xss/
    xss-python-flask.yaml     # Rules
    xss-python-flask.py       # Test file
  sqli/
    sqli-java-spring.yaml
    sqli-java-spring.java

Common Mistakes to Avoid

Mistake	Fix
Pattern too broad ($FUNC(...))	Add metavariable-regex or pattern-inside to constrain
Missing ellipsis in function bodies	Use ... in function/class bodies for flexible matching
Forgetting pattern-not for safe variants	Always consider: what does the SAFE code look like?
Wrong language ID	Check docs/ai/RULES_SYNTAX.md Section 9 for valid IDs
Taint without sanitizers	Always ask: what makes this data safe? Add sanitizers.
Hardcoding framework versions	Use ... for version-agnostic patterns
Not testing negatives	False positives destroy trust — test safe code paths