How AI Is Finding Weaknesses in Cryptographic Algorithms

Defined Learning Outcomes

After studying this material, you should be able to:

1. Explain core cryptographic concepts (digital signatures, symmetric ciphers, post-quantum cryptography)
2. Describe how AI models can identify mathematical flaws in cryptographic algorithms
3. Distinguish between implementation vulnerabilities and algorithmic vulnerabilities
4. Analyze the significance and limitations of the specific attacks discovered
5. Evaluate the broader implications of AI-assisted cryptanalysis for security research

---

Step 1: Foundation — What Is Cryptography and Why Does It Matter?

The Basic Problem Cryptography Solves

When you visit a website, two things must happen securely:

YOU ──────────────────────────── WEBSITE
     1. Verify it's really them  (Digital Signatures)
     2. Keep data private        (Symmetric Ciphers)

Two Core Cryptographic Tools

Why This Matters at Scale

• Email, banking, and all internet traffic depend on these systems
• Flaws could expose billions of users simultaneously
• Unlike a broken lock on one door, a broken cipher breaks every door using that design

---

Step 2: Two Types of Cryptographic Vulnerabilities

This is a critical distinction the article introduces early:

Type 1: Implementation Vulnerabilities

Algorithm (correct) ──► Programmer codes it incorrectly ──► Vulnerability

• The math is fine
• The code has errors
• Claude had already found these in major cryptographic libraries

Type 2: Algorithmic Vulnerabilities ← New territory in this article

Algorithm itself has a mathematical flaw ──► All correct implementations are vulnerable

• The code can be perfect
• The underlying math is broken
• Far more serious and harder to find

Key Insight: This article describes Claude moving from finding Type 1 to finding Type 2 vulnerabilities — a significant escalation in capability.

---

Step 3: Understanding the Post-Quantum Cryptography Context

Why "Post-Quantum" Matters

Current widely-used algorithms (RSA, ECDSA) rely on math problems that are:

• Hard for classical computers → secure today
• Potentially easy for quantum computers → vulnerable tomorrow

The NIST Standardization Process

NIST Issues Call for New Algorithms
           │
           ▼
Round 1: Many candidates submitted
           │
           ▼
Round 2: Experts review, weak ones eliminated
           │
           ▼
Round 3: Remaining candidates (including HAWK) ← We are here
           │
           ▼
Standardization: Algorithms deployed globally

The purpose of this process: Find weaknesses before deployment, not after.

Historical precedent: SIKE, a competing candidate, was completely broken in one hour on a laptop — during review, not after deployment.

---

Step 4: The HAWK Attack — What Happened and Why It Matters

What Is HAWK?

• A digital signature scheme designed to resist quantum computers
• Based on the Lattice Isomorphism Problem (a hard math problem)
• Had survived two rounds of expert human review over two years

Understanding "Key Strength"

Think of key strength like a combination lock:

More digits = More combinations = Harder to guess by brute force

HAWK-256 was thought to need 2^64 guesses to break
Claude showed it only needs 2^38 guesses

That reduction from 2⁶⁴ to 2³⁸ means:

The effective security was cut in half — equivalent to halving the number of digits on your combination lock

How Claude Found the Attack

The mathematical concept: A "nontrivial automorphism" — a hidden symmetry in the mathematical structure HAWK uses

Prior researchers knew:
"IF you find this symmetry, you can attack HAWK"
                    │
                    ▼
But they didn't know:
"Does this symmetry actually EXIST in HAWK's specific lattice?"
                    │
                    ▼
Claude discovered:
"YES — and here's how to exploit it"

The Multi-Agent Discovery Process

An interesting detail about how Claude found this:

Worker Agent 1: Investigates idea → Concludes "infeasible" → Stops
Worker Agent 2: Investigates same idea → Finds exploitation method → Continues
Both agents exchange messages → Reach consensus: "We found an attack"

This mirrors how human research teams work — different perspectives prevent premature rejection of valid ideas.

Practical Impact of the HAWK Finding

Bottom line: Doubling key sizes would restore security, but this eliminates HAWK's main advantages as a candidate. HAWK is not currently deployed, so no systems are at risk.

---

Step 5: The AES Attack — Understanding the Methodology

What Is AES?

• Advanced Encryption Standard — adopted 2001
• The most widely used symmetric cipher in the world
• AES-128 applies its "round function" 10 times
• Has received more scrutiny than almost any other algorithm

Why Study "Round-Reduced" Versions?

Full AES-128: 10 rounds ──► Too strong to attack directly
                                        │
                                        ▼
Researchers study: 7-round AES ──► Easier to attack
                                        │
                                        ▼
Goal: Understand attack techniques that might eventually scale up

This is like stress-testing a bridge by studying smaller versions — you learn about structural weaknesses without endangering real traffic.

The Attack Model (Chosen Plaintext)

Attacker says: "Encrypt THIS specific message for me"
Defender encrypts it with their secret key
Attacker sees the output
Attacker repeats this many times
Goal: Figure out the secret key from patterns in outputs

This is the standard assumption used in academic cipher research.

Claude's Innovation: The "Möbius Bridge"

The problem Claude solved:

Previous attacks required guessing one of 256 possible values at a critical step, then checking each guess against a lookup table:

Old approach:
Guess value 1 → Check table → Wrong
Guess value 2 → Check table → Wrong
...
Guess value 256 → Check table → Right!  (average: 128 guesses)

Claude's Möbius Bridge solution:

New approach:
Apply Möbius transform → Creates a "fingerprint" that works
regardless of which value is correct → Check table once → Right!

This eliminates the 256-guess enumeration entirely, making the attack 200-800× faster.

Why This Doesn't Break Real AES

7-round AES: BROKEN by Claude's attack
8-round AES: Not broken
9-round AES: Not broken
10-round AES (real AES): Not broken

Each additional round adds substantial security. The attack is academically significant but practically harmless to deployed systems.

---

Step 6: Additional Findings — Establishing a Pattern

The article briefly mentions other attacks Claude found, showing this isn't isolated:

Pattern: Claude consistently finds attacks on reduced-round versions, not full ciphers — but the trend is concerning as capabilities grow.

---

Step 7: The Broader Implications — What This Means

The Capability Trajectory

One year ago: LLMs couldn't perform basic cryptanalysis
Today: LLMs find flaws that escaped years of expert human review
Tomorrow: ???

The Verification Bottleneck Problem

A new asymmetry is emerging:

Claude discovers AES attack: ~1 week
Human researchers validate it: ~1 month (2 researchers)

Claude discovers HAWK attack: ~60 hours
Human researchers validate it: Much easier (end-to-end runnable)

Key Insight: AI can generate research faster than humans can verify it. This creates a new kind of bottleneck in security research.

Responsible Disclosure — How This Was Handled

The article describes a careful process:

1. Discovery → Internal validation
2. Consultation with academic experts
3. Advance sharing with US government and industry partners
4. Coordinated disclosure to NIST mailing list (for HAWK)
5. Public release of papers and findings

This mirrors standard security research ethics — notify affected parties before going public.

The Dual-Use Nature of This Research

Same capability that finds weaknesses in bad algorithms
                    │
                    ▼
Also finds weaknesses in good algorithms that are actually deployed

The article explicitly raises this concern: What if Claude finds a vulnerability in a cipher that IS currently protecting real systems?

---

Step 8: Synthesizing the Key Concepts

What Makes This Research Significant

The Three-Layer Security Model This Affects

Layer 1: Algorithm Design ← Claude now operates here
Layer 2: Implementation ← Claude already operated here
Layer 3: Deployment/Configuration ← Traditional security focus

What Hasn't Changed

• AES remains secure for real-world use
• Post-quantum standardization process is working as intended
• No production systems need to change based on these findings
• The cryptographic community's adversarial review model is validated, not undermined

---

Summary: Core Takeaways

1. Cryptographic security has two failure modes: bad implementations and bad algorithms. AI can now find both.

2. "Round-reduced" attacks are academically important even when they don't break full ciphers — they reveal structural weaknesses and improve our understanding.

3. The NIST standardization process exists precisely to catch problems like the HAWK weakness before deployment. Finding it now is the system working correctly.

4. AI's speed advantage in research creates a new challenge: human verification capacity may become the limiting factor in security research.

5. Responsible disclosure matters more as AI accelerates discovery — the gap between "found" and "exploited by bad actors" may shrink.

6. The trajectory is the most important signal: going from zero cryptanalytic capability to expert-level in one year suggests this capability will continue growing.

---

Self-Check Questions

1. What is the difference between breaking an algorithm's implementation versus breaking the algorithm itself? Why does the distinction matter?

2. Why does NIST run a multi-year public review process for cryptographic candidates? How does the HAWK finding validate this approach?

3. Explain in plain language why attacking 7-round AES is academically valuable even though it doesn't break real AES.

4. What is the "verification bottleneck" problem, and why does it matter for the future of security research?

5. If Claude found a critical vulnerability in a widely-deployed cipher tomorrow, what considerations would govern how that finding should be handled?