Anthropic's Claude Mythos Preview Accelerates Cryptanalytic Attack on Reduced-Round AES

Anthropic's frontier AI model, Claude Mythos Preview, has achieved a significant breakthrough in cryptanalysis, demonstrating an ability to accelerate a known attack on a reduced-round variant of the Advanced Encryption Standard (AES) by 200 to 800 times. This development, announced by Anthropic on July 28, 2026, highlights the rapidly advancing capabilities of artificial intelligence in uncovering mathematical weaknesses within cryptographic algorithms. While the findings do not pose an immediate threat to the full, production-grade AES standard widely used today, they underscore AI's emergent capacity for autonomous cryptographic research and raise critical questions about the future of digital security.

1. AI Breakthrough in Cryptanalysis

Claude Mythos Preview, a powerful, unreleased AI model developed by Anthropic, has showcased its advanced reasoning and coding capabilities by identifying an improved method to attack a simplified version of AES. The model's success marks a pivotal moment for AI-driven cryptanalysis, moving beyond merely finding implementation bugs to discovering fundamental mathematical flaws within cryptographic designs themselves. This capability was developed largely autonomously, with human researchers primarily guiding the project and verifying the results.

The specific target of this accelerated attack was a 7-round variant of AES-128, a common research target for cryptanalysts, rather than the full 10-round AES-128 standard employed in real-world systems. Cryptographers routinely study reduced-round versions of algorithms to understand their security margins and test potential attack methodologies. Anthropic has explicitly stated that these findings do not compromise any currently deployed encryption systems, as the full AES standard remains robust and unaffected by this particular discovery.

2. The "Möbius Bridge" and Attack Details

The Advanced Encryption Standard (AES) is a symmetric block cipher adopted by the U.S. government in 2001 and is the most widely used symmetric encryption standard globally, securing everything from online banking to encrypted communications. Full AES-128 operates over 10 rounds of cryptographic transformations. The variant attacked by Claude Mythos Preview used only 7 of these 10 rounds.

Mythos Preview’s key innovation was the discovery of a mathematical shortcut dubbed the "Möbius Bridge". This shortcut was applied to an existing "meet-in-the-middle" attack, a common cryptanalytic technique that involves working both forward from the plaintext and backward from the ciphertext to find a common intermediate state. The "Möbius Bridge" effectively eliminated a 256-way guessing step previously required in such attacks, leading to the substantial speed improvement.

Despite the dramatic acceleration, the improved attack on 7-round AES-128 still demands an impractical amount of computational resources and data in real-world scenarios. For instance, it would require an attacker to obtain approximately 2^105 chosen plaintexts and perform around 2^89 cipher operations, conditions that are currently unachievable with existing technology.

3. Autonomous Discovery and Human Oversight

The development of this advanced cryptanalytic technique by Claude Mythos Preview was largely autonomous. Anthropic researchers created a scaffold that allowed the AI to generate hypotheses, conduct experiments to validate or refute them, and then design an improved attack. Human involvement was primarily limited to project management, providing initial prompts, and subsequently verifying the model’s complex outputs. This process demonstrated the AI's ability to engage in original cryptographic research at a level comparable to expert human cryptanalysts.

Anthropic noted that the discovery process for the AES attack took approximately one week for the AI model. However, the subsequent validation of the results by two human researchers extended for nearly a month, highlighting the challenges of verifying AI-generated research. The estimated cost for the API usage associated with each major cryptanalytic result (including another significant finding on the post-quantum digital signature scheme HAWK) was approximately $100,000.

In parallel with the AES discovery, Claude Mythos Preview also uncovered a substantial weakness in HAWK, a digital signature scheme under consideration by the National Institute of Standards and Technology (NIST) for post-quantum cryptography. This separate finding effectively halved HAWK's key strength, further illustrating the AI model's formidable capabilities in cryptanalysis. Anthropic followed responsible disclosure practices, informing HAWK's designers and coordinating public release with NIST.

4. Implications for Cryptographic Research and AI Safety

This research from Anthropic marks a significant evolution in the role of AI in cybersecurity. Previously, AI models primarily assisted in finding implementation flaws in software. The current work shows a qualitative leap, demonstrating AI's ability to uncover mathematical weaknesses in the underlying algorithms themselves, a domain traditionally requiring deep human expertise.

While the immediate impact on deployed systems is minimal, the long-term implications for cryptographic research are substantial. The rapid pace at which AI models can conduct complex cryptanalytic investigations suggests a potential future where AI could outpace human experts in discovering vulnerabilities. This raises critical questions for researchers, companies, and governments about how to manage and respond to AI systems that could potentially uncover flaws in cryptographic systems protecting critical infrastructure.

Anthropic emphasizes the importance of a proactive approach to AI safety and responsible development. The company plans to host an academic workshop to facilitate discussions among researchers on the role of language models in security and cryptography research, aiming to address these complex challenges collaboratively. The ongoing evaluation process for cryptographic candidates, such as those in NIST's post-quantum competition, benefits from such rigorous analysis, even if conducted by AI.

5. Understanding Reduced-Round Cryptanalysis

The practice of cryptanalyzing "reduced-round" versions of ciphers is a standard and valuable technique in cryptographic research. Full cryptographic algorithms, like AES, are designed with a sufficient number of rounds to ensure security against all known attacks. By removing rounds, cryptographers create simpler targets that allow them to:

1. Probe Security Margins: Determine how many rounds are truly necessary for the cipher's security. 2. Develop New Techniques: Experiment with novel cryptanalytic methods on easier targets before attempting them on the full cipher. 3. Validate Design Principles: Understand if the cipher's design principles hold up even under weakened conditions.

The fact that Claude Mythos Preview succeeded against a 7-round variant of AES-128 is a research achievement that contributes to this understanding. However, it does not imply that the full AES, with its 10, 12, or 14 rounds (depending on key size), is vulnerable. The computational and data requirements for this specific attack remain astronomically high, rendering it impractical for real-world exploitation against current AES implementations. This research serves as a forward-looking indicator of AI's potential in cryptanalysis, rather than an immediate security concern for deployed systems.

Frequently Asked Questions

What is Claude Mythos Preview?

Claude Mythos Preview is a frontier AI model developed by Anthropic, not yet publicly released, known for its advanced capabilities in cybersecurity, autonomous coding, and complex reasoning.

Does this mean AES is broken?

No, the full Advanced Encryption Standard (AES) used in production systems today is not broken or compromised by this research. The attack specifically targets a reduced-round, 7-round variant of AES-128 used for research purposes.

What is "reduced-round AES"?

"Reduced-round AES" refers to a modified version of the AES algorithm that uses fewer cryptographic rounds than the full standard. Cryptographers study these weakened variants to understand an algorithm's security margins and test new attack techniques.

What are the practical implications of this research?

This research demonstrates AI's growing ability to conduct autonomous cryptographic research and discover mathematical flaws in algorithms, rather than just implementation bugs. While not affecting current systems, it highlights the need for ongoing vigilance and research into AI's role in cryptanalysis and cybersecurity.

Sources

* Original X post by @AnthropicAI * Anthropic: Discovering cryptographic weaknesses with Claude * CyberScoop: Anthropic's Claude Mythos finds weaknesses in encryption algorithms * The Decoder: Anthropic says its Mythos model found vulnerabilities in cryptographic algorithms that secure the internet * CSO Online: Mythos takes its first shot at post-quantum cryptography * Rescana: AI-Powered Cryptanalysis: Claude Mythos Uncovers HAWK Post-Quantum Weakness and Accelerated 7-Round AES Attack * The Hacker News: Claude AI Just Cracked a Post-Quantum Test Scheme and Found a Faster 7-Round AES Attack * TNW: Claude found mathematical flaws in two cryptographic algorithms that years of expert review missed * Matthew Green: Some thoughts about Anthropic's new cryptanalysis results * Security Affairs: Claude Mythos Shows AI Can Outpace Human Cryptography Research * RuntimeWire: Anthropic says Claude improved attacks on HAWK and reduced-round AES