OWASP Top 10 2025: A06-A10 & LLM Top 10

Completing our OWASP coverage with the remaining five web vulnerabilities plus the LLM-specific risks that are increasingly tested in AI security roles.

A06:2025 - Insecure Design

Rank: #6 (was #4 in 2021)

Flaws in the design phase that cannot be fixed by perfect implementation.

Design Flaws vs Implementation Bugs

Category	Example	Fix
Insecure Design	No rate limiting designed	Redesign with rate limiting
Implementation Bug	Rate limiter bypass	Fix the code

Interview Scenario

Q: "Design a password reset flow. What security considerations matter?"

Strong Answer:

• Token generation: Cryptographically random, 32+ bytes
• Token expiration: Short-lived (15-60 minutes)
• Single use: Invalidate after use
• Rate limiting: Prevent enumeration
• No user info in URL: Token only
• Secure delivery: Email, not SMS
• Account lockout: After N failed attempts

A07:2025 - Authentication and Identification Failures

Rank: #7 (unchanged)

Weaknesses in authentication mechanisms and session management.

Common Failures

Failure	Risk	Mitigation
Weak passwords	Credential stuffing	Password policies, breach checks
Missing MFA	Account takeover	Require MFA for sensitive actions
Session fixation	Session hijacking	Regenerate session on auth
Credential exposure	Mass compromise	Secure storage, rotation

Interview Code Review

# Find the vulnerabilities
def login(username, password):
    user = db.query(f"SELECT * FROM users WHERE username='{username}'")
    if user and user.password == password:  # Issue 1: SQL injection
        session['user_id'] = user.id        # Issue 2: Plaintext comparison
        return redirect('/dashboard')       # Issue 3: No session regeneration
    return "Invalid credentials"            # Issue 4: Generic error is OK

A08:2025 - Software and Data Integrity Failures

Rank: #8

Failures to verify integrity of code, data, or configurations.

Attack Scenarios

Scenario	Description	Defense
Unsigned updates	Malicious update pushed	Code signing, verification
Insecure deserialization	RCE via object injection	Avoid deserialization of untrusted data
CI/CD tampering	Build pipeline compromise	Signed commits, immutable artifacts

# Vulnerable deserialization
import pickle

def process_data(serialized):
    return pickle.loads(serialized)  # RCE if attacker controls input

# Safer alternative
import json

def process_data(serialized):
    return json.loads(serialized)  # Only data, no code execution

A09:2025 - Security Logging and Alerting Failures

Rank: #9 (expanded from "Insufficient Logging")

Logs without alerting and action provide limited value.

What to Log

Event	Priority	Retention
Login attempts (success/fail)	High	90 days
Access control failures	High	90 days
Input validation failures	Medium	30 days
System errors	Medium	30 days

Interview Question

Q: "What would you log for a financial transaction API?"

Answer:

{
  "timestamp": "2026-01-05T10:30:00Z",
  "event_type": "TRANSACTION",
  "user_id": "hashed_user_123",
  "action": "TRANSFER",
  "amount": 1000.00,
  "source_ip": "192.168.1.1",
  "user_agent": "Mozilla/5.0...",
  "request_id": "uuid-xxx",
  "result": "SUCCESS",
  "latency_ms": 45
}

Key: Log enough for investigation but never log sensitive data (passwords, full card numbers, PII).

A10:2025 - Mishandling of Exceptional Conditions

Rank: #10 (NEW)

When applications don't handle unexpected situations properly.

Common Issues

Issue	Example	Impact
Unhandled exceptions	Stack trace in response	Information disclosure
Resource exhaustion	No timeout on operations	DoS
Race conditions	TOCTOU vulnerabilities	Privilege escalation
Fail-open designs	Allow on error	Security bypass

# Fail-open (WRONG)
def check_authorization(user, resource):
    try:
        return authorization_service.check(user, resource)
    except Exception:
        return True  # DANGER: Fail-open

# Fail-closed (CORRECT)
def check_authorization(user, resource):
    try:
        return authorization_service.check(user, resource)
    except Exception as e:
        log.error(f"AuthZ check failed: {e}")
        return False  # Fail-closed is secure

---

OWASP LLM Top 10 2025

For AI security roles, know the LLM-specific vulnerabilities:

Rank	Vulnerability	Description
LLM01	Prompt Injection	Malicious prompts manipulating LLM behavior
LLM02	Sensitive Information Disclosure	LLM revealing training data or secrets
LLM03	Supply Chain	Compromised models, plugins, or training data
LLM05	Improper Output Handling	XSS/injection via LLM output
LLM06	Excessive Agency	LLM with too many permissions
LLM07	System Prompt Leakage	Exposing system instructions
LLM08	Vector and Embedding Weaknesses	RAG poisoning attacks

Interview Question (AI Security Role)

Q: "How would you prevent prompt injection attacks?"

Answer:

1. Input sanitization and filtering
2. Output encoding before rendering
3. Privilege separation (LLM can't execute code directly)
4. Prompt/completion separation
5. Rate limiting and monitoring
6. Human-in-the-loop for sensitive actions

For AI Security Roles: The LLM Top 10 is now standard interview material. Expect detailed questions on prompt injection and RAG security.

Next, we'll cover cryptography essentials for security interviews. :::