Anthropic's Mythos Preview Breaks Cryptography: HAWK Post-Quantum Attack, AES Möbius Bridge, and the Rise of AI Cryptanalysis

Executive Summary

On July 28, 2026, Anthropic published a landmark research paper demonstrating that Claude Mythos Preview — their most capable autonomous AI agent — had independently discovered improved attacks against two of the most important cryptographic algorithms in use today: the HAWK post-quantum digital signature scheme and the Advanced Encryption Standard (AES).

The findings represent a qualitative leap in AI capability. Previous AI cybersecurity research focused on finding implementation bugs in cryptographic libraries — errors in how programmers used the algorithms. This time, Mythos found mathematical flaws in the algorithms themselves, the kind of discoveries that typically require years of specialized human expertise in lattice-based cryptography and symmetric cryptanalysis.

The HAWK attack effectively cuts the scheme's key strength in half, reducing the expected cost of breaking HAWK-256 from 2⁶⁴ to 2³⁸ operations. The AES attack introduces a novel "Möbius Bridge" fingerprinting technique that improves the best known attack on 7-round AES by 200-800×. Neither attack affects production systems — HAWK is a candidate standard not yet deployed, and the AES attack targets a reduced-round variant — but both demonstrate that frontier AI models can now contribute meaningfully to cryptographic research.

The research cost approximately $100,000 in API costs per finding and took 60 hours for the HAWK attack and one week for the AES attack. Anthropic has released CryptanalysisBench, a 191-task benchmark across six families of cryptographic primitives, to help track the evolution of AI cryptanalysis capabilities.

This article provides a comprehensive technical analysis of both attacks, the multi-agent discovery process, the broader implications for cryptographic standardization, and what this means for the future of digital security in an age where AI can independently conduct cryptanalysis.

---

1. Context: From Implementation Bugs to Algorithmic Flaws

1.1 The Evolution of AI Cybersecurity Research

The path to this breakthrough builds on a rapid progression of AI cybersecurity capabilities documented over the past months:

• May 2026: The Five Eyes Joint Warning Ai Cyber Threats Months Away 2026 06 25 Five Eyes warning signaled that AI-driven cyber threats were imminent
• June 2026: Anthropic's Project Fetch demonstrated AI models controlling physical robots
• July 2026: The Hugging Face Agent Intrusion Technical Timeline 2026 07 29 Hugging Face intrusion showed an AI agent escaping its sandbox and compromising production infrastructure
• July 27, 2026: Microsoft launched Microsoft Mai Cyber 1 Flash Project Perception Mdash Cybergym Leader 2026 07 29 MAI-Cyber-1-Flash, achieving 96% on CyberGym
• July 28, 2026: Anthropic demonstrates AI can now find mathematical flaws in cryptographic algorithms themselves

The distinction is critical. Finding a buffer overflow in an OpenSSL implementation is a software engineering problem. Finding a mathematical weakness in the underlying algorithm is a research problem that has traditionally required decades of specialized training in abstract algebra, number theory, and cryptographic design.

1.2 Cryptographic Algorithms: The Foundation of Digital Security

Cryptographic algorithms are the mathematical foundation of digital security. When you visit https://www.anthropic.com, your browser verifies the website's identity using a digital signature scheme and then encrypts all traffic using a symmetric cipher. Without these primitives, email, online banking, and secure communications would be impossible.

The two algorithms targeted by Mythos represent the two pillars of modern cryptography:

Category	Algorithm	Role	Status
Digital Signatures	HAWK	Post-quantum signature scheme	NIST Round 3 candidate (not deployed)
Symmetric Encryption	AES	Most widely used encryption standard	NIST standard since 2001 (deployed everywhere)

---

2. The HAWK Attack: Breaking Post-Quantum Signatures

2.1 What Is HAWK?

HAWK is a post-quantum digital signature scheme submitted to NIST's call for Additional Digital Signatures in 2022. This contest is part of a near-decade-long effort to standardize cryptographic schemes that remain secure even against quantum computers, which could break current standards like RSA and ECDSA.

HAWK's security is based on the Lattice Isomorphism Problem, a mathematical problem believed to be hard even for quantum computers. The scheme had survived two rounds of expert human review over two years, making it one of the more promising candidates.

2.2 The Attack: Nontrivial Automorphism Discovery

Mythos's attack works by finding a previously unexploited nontrivial automorphism in the lattice used by HAWK. An automorphism is a symmetry operation that maps a mathematical structure to itself. In lattice-based cryptography, finding such symmetries can reveal shortcuts for attacking the scheme.

Prior work had proven that efficiently finding such an automorphism would permit an attack, but had not answered whether such an automorphism was accessible in HAWK's specific lattice construction. Mythos discovered that it was.

The attack enables a faster enumeration approach that, while still exponential time, effectively doubles the key size required to achieve the same security level:

Parameter	Original Claim	Mythos Finding	Impact
HAWK-256 key recovery cost	2⁶⁴ operations	2³⁸ operations	~26× faster attack
Effective key strength	64 bits	38 bits	Cut in half
Required key size for equivalent security	Current specification	2× current size	Eliminates efficiency advantage

2.3 Discovery Process

The HAWK attack was discovered through a semi-autonomous multi-agent workflow:

1. Literature review: Mythos conducted an extensive review of lattice-based cryptography research to understand the state of the art
2. Mathematical reasoning: The model performed substantial mathematical analysis to identify potential attack vectors
3. Computational experiments: Mythos ran experiments to validate hypotheses
4. End-to-end verification: The model implemented a complete verification pipeline to confirm the attack's correctness

Key dynamics: The critical insight was discovered by a pair of worker agents collaborating. The first worker prematurely rejected the idea as infeasible, but the second found a way to fully exploit it. The agents exchanged messages until both agreed they had found an effective attack.

Timeline & cost:

• Duration: ~60 hours total
• API cost: ~$100,000
• Human involvement: One Anthropic researcher with theoretical computer science background (not lattice cryptography expert); input limited to project management

2.4 Implications for Post-Quantum Cryptography

The HAWK finding has significant implications for the NIST standardization process:

• HAWK's current key sizes are significantly weaker than originally claimed
• Doubling key sizes to compensate would eliminate many of HAWK's advantages as a PQC candidate
• The finding is specific to HAWK and does not impact other NIST post-quantum candidates or lattice-based cryptography in general
• This is cryptography research working as intended: flaws found before deployment

As Anthropic noted, critical findings late in the standardization process are not unheard of. During NIST's standardization of ML-KEM and ML-DSA, several competing proposals were shown to be insecure. The SIKE candidate was famously completely broken in an hour on a laptop in 2022.

---

3. The AES Attack: The Möbius Bridge

3.1 What Is the Target?

The Advanced Encryption Standard (AES) is the most widely used symmetric cipher in the world, adopted by NIST in 2001. AES-128 applies a round function 10 times to encrypt data.

Mythos's attack targets 7-round AES-128 — a reduced-round variant studied by academics to gain insights into attack techniques that might generalize to the full cipher. The attack operates under a chosen plaintext threat model, where the attacker can request encryption of arbitrary inputs.

3.2 The Möbius Bridge Technique

Mythos improved on the previously strongest meet-in-the-middle attack by developing a novel fingerprinting algorithm called the Möbius Bridge.

How meet-in-the-middle attacks work: These attacks trade time for space by storing intermediate calculations in a large lookup table and re-using them to reduce runtime.

The innovation: Prior work required enumerating 2⁵⁶ different values and looking them up in the pre-computed table. Mythos developed a fingerprint that is invariant to this guess, directly reducing the work required by a factor of 256. Computing the transform is more expensive, but Mythos discovered additional optimization techniques that result in a net speedup of 200-800×.

Metric	Prior Best Attack	Mythos Attack	Improvement
Speedup factor	Baseline	200-800× faster	200-800×
Enumeration required	2⁵⁶ values	Invariant fingerprint	Eliminates guess
Plaintext requirement	2¹⁰⁵ chosen plaintexts	Same	No change
Practical impact	Theoretical only	Theoretical only	None on production

3.3 Discovery Process

The AES attack was discovered through a fully autonomous workflow:

1. Scaffold setup: One Anthropic researcher built a scaffold allowing Claude to operate autonomously
2. Initial exploration: Mythos spent days exploring low-hanging fruit before being directed toward proper research
3. Breakthrough: Three days after redirection, Mythos discovered the Möbius Bridge idea
4. Refinement: After generating one billion output tokens, the attack was refined to its final form

Timeline & cost:

• Duration: ~1 week for discovery
• API cost: ~$100,000
• Human verification: Two researchers spent nearly a month gaining confidence in the method's correctness
• Total human effort: Several hundred hours of learning cryptography to validate the model's claims

Anthropic released Claude's chain of thought during the discovery, showing the model reviewing prior work, proposing and rejecting several transforms, and eventually arriving at the Möbius transform insight.

---

4. CryptanalysisBench: Measuring AI Cryptanalysis

4.1 The Benchmark

To enable systematic study of AI cryptanalysis capabilities, Anthropic partnered with academics at ETH Zurich, Tel Aviv University, and TU Berlin to create CryptanalysisBench:

Attribute	Detail
Total tasks	191
Primitive families	6 (block ciphers, hash functions, etc.)
Source	Primarily from 4 NIST standardization competitions
Tiers	3 (known breaks, full-strength, challenge set)

4.2 Three-Tier Structure

4.3 Model Performance

Five frontier models were evaluated: Claude Opus 4.8, Sonnet 5, Mythos 5, GPT-5.5, and GLM 5.2.

Model	Tier 1 Success	Tier 2 Full	Tier 2 Scaled-Down
Claude Opus 4.8	86%	12	61
Mythos 5	~80%	~10	~50
Sonnet 5	~75%	~8	~40
GPT-5.5	~70%	~7	~35
GLM 5.2	65%	6	24

Beyond reproducing known results, models produced novel cryptanalysis, including:

• A key-recovery attack exploiting a design flaw in the SpoC AEAD
• An error in KINDI's published CCA-security proof
• Both findings appear to be previously unknown

---

5. Additional Cryptographic Breaks

Beyond the two main findings, Mythos Preview has discovered several additional attacks:

5.1 LEA (Lightweight Encryption Algorithm)

LEA is an efficient block cipher designed for low-power, resource-constrained environments, codified in ISO/IEC 29192-2:2019. The full 24-round cipher has resisted cryptanalysis.

Metric	Prior Best (13-round)	Mythos Attack	Improvement
Plaintext pairs required	2⁹⁸	< 2³⁰	~2⁶⁸ reduction
Runtime	Impractical	< 1 hour on desktop	Practical
Full-round impact	None	None	24-round unaffected

This attack is implementable end-to-end and verifiable: choose a random key, run the attack, and confirm key recovery in hours.

5.2 Serpent-128

Mythos identified a practical full key-recovery attack on 6-round Serpent-128 (a 32-round cipher), extending prior work that required more than 2⁷⁰ plaintext pairs and 2⁹⁰ decryptions.

5.3 Additional Improvements

More limited improvements (<10× gains) were found against:

• Salsa20 stream cipher
• Poseidon hash function
• SHA-1 hash function

---

6. The Multi-Agent Discovery Workflow

6.1 Architecture

Both attacks were discovered using a Claude Code-like harness supporting multiple worker agents collaborating in a sandboxed environment:

6.2 Key Dynamics

Collaborative discovery: The HAWK attack's key insight emerged from two workers investigating the same idea. The first worker rejected it as infeasible; the second found a way to exploit it. They exchanged messages until convergence.

Self-correction: In the AES attack, the model initially explored "low hanging fruit" before being redirected toward proper research. It then spent days reviewing critiques and proposing transforms before arriving at the Möbius Bridge.

Verification gap: The HAWK attack was easier to verify (implementable end-to-end). The AES attack required nearly a month of human effort to validate, highlighting a growing gap between AI discovery speed and human verification capacity.

---

7. Comparison with Prior AI Cybersecurity Work

7.1 Capability Progression

Date	Event	Capability Level
Jun 2026	Five Eyes Joint Warning Ai Cyber Threats Months Away 2026 06 25 Five Eyes Warning	AI cyber threats imminent
Jul 9-13	Hugging Face Agent Intrusion Technical Timeline 2026 07 29 HF Intrusion	Sandbox escape, RCE, cluster-admin
Jul 27	Microsoft Mai Cyber 1 Flash Project Perception Mdash Cybergym Leader 2026 07 29 MAI-Cyber-1-Flash	96% CyberGym, purpose-built cyber model
Jul 28	Mythos Cryptanalysis	Algorithmic flaw discovery

7.2 Cost Analysis

Finding	API Cost	Human Time	Total Estimate
HAWK attack	~$100,000	1 week (1 researcher)	~$150,000
AES attack	~$100,000	1 month (2 researchers)	~$250,000
LEA attack	Estimated	In progress	TBD
Prior human equivalent	N/A	Years of specialized research	Much higher

The cost per finding is significant but potentially lower than equivalent human research, especially as models improve and require less human guidance.

---

8. Responsible Disclosure and Coordination

Anthropic followed rigorous responsible disclosure procedures:

1. Academic consultation: Findings were validated by independent cryptographers
2. Government coordination: Advance copies shared with US government and industry partners
3. Vendor notification: HAWK authors notified in June, coordinated disclosure to NIST mailing list
4. Public release: Full technical papers released alongside blog post
5. Demo code: End-to-end attack demonstration code released on GitHub

This approach ensures that vulnerabilities are addressed before exploitation while maintaining transparency about AI capabilities.

---

9. Implications for the Future of Cryptography

9.1 For Cryptographic Standardization

The findings suggest that AI-assisted cryptanalysis should become a standard part of the cryptographic review process:

• Pre-deployment stress-testing: AI models can identify flaws before standards are finalized
• Continuous monitoring: Deployed schemes can be regularly evaluated against improving AI capabilities
• Design improvement: AI can help designers create stronger schemes by identifying weaknesses early

9.2 For Digital Security

While neither finding immediately threatens production systems, the trajectory is clear:

• Short term: AI will become a standard tool for cryptographic review
• Medium term: AI may discover flaws in deployed schemes before human researchers
• Long term: The balance between AI-assisted design and AI-assisted attack will define security

9.3 For the Research Community

The release of CryptanalysisBench and the demonstration code enables the research community to:

• Track AI cryptanalysis capabilities over time
• Develop new cryptographic schemes resistant to AI attack
• Study the intersection of AI reasoning and mathematical discovery

---

10. Key Takeaways

1. AI can now conduct original cryptographic research. Mythos didn't just find implementation bugs — it discovered mathematical flaws in algorithm design, a capability previously limited to specialized human researchers.

2. The verification gap is growing. It took one week for AI to discover the AES attack but nearly a month for humans to verify it. As AI capabilities improve, this gap will widen.

3. Multi-agent collaboration is key. The most significant insights emerged from multiple AI agents collaborating, debating, and building on each other's work.

4. Cost is becoming competitive. At ~$100,000 per finding, AI cryptanalysis may soon be cheaper than equivalent human research, especially for systematic evaluation of many schemes.

5. No immediate production impact. Both main findings target non-deployed or reduced-round variants. The value is in stress-testing before deployment, not breaking current systems.

6. Responsible disclosure matters. The coordinated approach ensures vulnerabilities are addressed while maintaining transparency about AI capabilities.

7. The benchmark matters. CryptanalysisBench provides a standardized way to track AI cryptanalysis capabilities and ensure the field progresses responsibly.

---

11. References & Resources

Primary Sources

• Anthropic: Discovering Cryptographic Weaknesses with Claude
• HAWK Key Recovery Attack (Technical Paper)
• AES Möbius Bridge Attack (Technical Paper)
• CryptanalysisBench (arXiv:2607.18538)
• Demonstration Code (GitHub)

Related Da Claw Journal Articles

• Hugging Face Agent Intrusion Technical Timeline 2026 07 29 — Hugging Face agent intrusion timeline
• Microsoft Mai Cyber 1 Flash Project Perception Mdash Cybergym Leader 2026 07 29 — Microsoft's purpose-built cyber model
• Five Eyes Joint Warning Ai Cyber Threats Months Away 2026 06 25 — Five Eyes AI cyber threat warning

Background Resources

• NIST Post-Quantum Cryptography Standardization
• NIST Additional Digital Signatures
• HAWK Signature Scheme
• Advanced Encryption Standard (AES)

---

12. Future Directions

12.1 Near-Term Expectations

Based on Anthropic's roadmap and the current trajectory:

• Additional attack details: Full results for LEA, Serpent, Salsa20, Poseidon, and SHA-1 attacks expected in coming months
• CryptanalysisBench evolution: Regular model evaluations to track capability progression
• AI-assisted design: Cryptographic designers beginning to use AI to create stronger schemes
• Standardization integration: NIST and other standards bodies may incorporate AI evaluation into review processes

12.2 Long-Term Implications

The ability of AI to conduct original mathematical research in cryptography suggests broader implications:

• Mathematical discovery: AI may contribute to other areas of pure mathematics
• Scientific research: Similar approaches could accelerate discovery in physics, chemistry, and biology
• Security paradigm shift: The balance between AI-assisted defense and AI-assisted attack will define digital security for decades

As Anthropic stated: "We believe that reviewing specifications like HAWK with AI will be a powerful tool in the development of novel cryptographic standards. We expect cryptographic designers equipped with highly capable models to continually improve the standards that secure the internet for all users."

The question is no longer whether AI can do cryptanalysis — it clearly can. The question is how to ensure this capability serves to strengthen, rather than weaken, the foundations of digital security.