.plan.md

AdvisoryVerifier Enhancement Plan

Target: hooks/learning-loop/post_tool_learning.py Scope: Fix blockers + expand pattern coverage

---

BLOCKER 1: Comment False Positives

Problem: Lines like # eval() is dangerous trigger warnings

Solution: Add comment filtering in _get_added_lines() or analyze_edit()

def _is_comment_line(line: str) -> bool:
    """Check if line is a comment (Python, JS, shell, etc.)"""
    stripped = line.strip()
    return (
        stripped.startswith('#') or      # Python, shell
        stripped.startswith('//') or     # JS, C, Go
        stripped.startswith('*') or      # Multi-line comment body
        stripped.startswith('/*') or     # C-style start
        stripped.startswith('"""') or    # Docstring
        stripped.startswith("'''")       # Docstring
    )

Integration point: Filter in _get_added_lines() before returning

---

BLOCKER 2: Password False Negatives

Problem: print("password: admin") not detected

Current pattern:

["\']?password["\']?\s*[:=]\s*["\'][^"\']+["\']

Enhanced patterns (add alternatives):

# Pattern 1: Assignment style (existing)
r'["\']?password["\']?\s*[:=]\s*["\'][^"\']+["\']'

# Pattern 2: Colon inside string (password: value)
r'["\']password:\s*[^"\']{3,}["\']'

# Pattern 3: Password as function arg with value
r'password\s*=\s*["\'][^"\']+["\']'

# Pattern 4: JSON-style "password": "value"
r'"password"\s*:\s*"[^"]+"'

Also add secret detection for:

• secret, token, credential, auth

---

EXPANSION: New Pattern Categories

High Priority (Common Vulnerabilities)

1. Insecure Deserialization

   (r'pickle\.loads?\s*\(', 'pickle.load/loads - insecure deserialization risk')
   (r'yaml\.load\s*\([^)]*\)', 'yaml.load without SafeLoader - potential code execution')
   (r'marshal\.loads?\s*\(', 'marshal.load - insecure deserialization')

2. Weak Cryptography

   (r'hashlib\.md5\s*\(', 'MD5 hash - cryptographically weak')
   (r'hashlib\.sha1\s*\(', 'SHA1 hash - cryptographically weak')
   (r'random\.\w+\s*\(', 'random module - not cryptographically secure')

3. Command Injection

   (r'os\.system\s*\(', 'os.system - potential command injection')
   (r'os\.popen\s*\(', 'os.popen - potential command injection')

4. Path Traversal

   (r'open\s*\([^)]*\+[^)]*\)', 'File open with string concatenation - potential path traversal')
   (r'\.\./', 'Path traversal pattern detected')

Medium Priority

5. Network Security

   (r'verify\s*=\s*False', 'SSL verification disabled')
   (r'requests\.get\s*\([^)]*(?!timeout)[^)]*\)', 'requests without timeout')

6. Secrets in Other Formats

   (r'(SECRET|TOKEN|KEY)\s*=\s*["\'][^"\']+["\']', 'Hardcoded secret/token/key')
   (r'Bearer\s+[A-Za-z0-9_-]+', 'Hardcoded bearer token')

7. XML Security

   (r'xml\.etree\.ElementTree\.parse', 'XML parse without DTD protection - XXE risk')

---

Implementation Structure

RISKY_PATTERNS = {
    'code_injection': [
        # eval, exec, subprocess shell=True
    ],
    'secrets': [
        # password, api_key, token, secret (enhanced)
    ],
    'deserialization': [
        # pickle, yaml, marshal
    ],
    'cryptography': [
        # MD5, SHA1, random
    ],
    'command_injection': [
        # os.system, os.popen
    ],
    'path_traversal': [
        # ../, string concat in open()
    ],
    'network': [
        # verify=False, no timeout
    ],
    'file_operations': [
        # rm -rf, chmod 777, /etc/ writes
    ]
}

---

Testing Requirements

After implementation:

1. Re-run existing 9 pattern tests (regression)
2. Test comment filtering (no false positives)
3. Test enhanced password detection (catches more cases)
4. Test each new pattern category
5. Verify no blocking behavior (advisory only)

---

Teammate Assignment

Teammate	Task
1	Fix comment false positives in _get_added_lines()
2	Enhance secret detection (password, token, key, secret)
3	Add new pattern categories (deserialization, crypto, command injection, etc.)

All teammates work on: hooks/learning-loop/post_tool_learning.py