Discovering Cryptographic Weaknesses with Claude
Anthropic's Claude has evolved from finding implementation bugs to discovering mathematical flaws in cryptographic algorithms, impressively weakening a post-quantum candidate (HAWK) and improving attacks on reduced-round AES. This demonstration of advanced AI capabilities in cryptanalysis hints at a future where models play a critical, albeit expensive, role in securing digital systems. The community grapples with the implications of AI-driven vulnerability discovery and the high computational costs involved.
The Lowdown
Anthropic's Claude Mythos Preview has achieved a significant milestone, moving beyond identifying software implementation vulnerabilities to uncovering fundamental mathematical weaknesses within cryptographic algorithms themselves. While these findings don't currently impact deployed production systems, they highlight the burgeoning capabilities of frontier AI models in highly specialized research domains.
- HAWK Attack: Claude semi-autonomously discovered an improved attack against HAWK, a digital signature scheme under consideration by NIST for post-quantum cryptography. This attack, achieved in just 60 hours at an estimated API cost of $100,000, effectively halves HAWK's key strength by exploiting a previously unknown symmetry.
- AES Attack: The AI autonomously developed a novel "Möbius Bridge" technique to improve attacks on a 7-round variant of AES-128, making it 200-800 times faster than prior methods. Interestingly, Claude initially expressed reluctance to tackle this problem, requiring researcher encouragement.
- Practical Implications: Both attacks target candidate schemes or reduced-round versions, meaning no immediate real-world systems are affected. However, they serve as crucial stress tests, pushing cryptographic boundaries and demonstrating AI's potential in pre-deployment vetting and security research.
- Responsible Disclosure & Benchmarking: Anthropic followed responsible disclosure protocols, consulting with academics and sharing findings with government and industry partners. They also released CryptanalysisBench, a benchmark to facilitate further research into LLM cryptanalysis capabilities.
- Further Discoveries: Claude has also identified practical key-recovery attacks on 13-round LEA and 6-round Serpent-128, along with minor improvements on Salsa20, Poseidon, and SHA-1.
This research underscores that AI models are becoming potent cryptanalytic tools, capable of expert-level mathematical discoveries. As these capabilities advance, the challenge for human researchers will increasingly shift towards validating AI-generated findings and navigating the complex ethical and practical considerations of AI-driven vulnerability discovery in critical systems.
The Gossip
Costly Cryptanalysis & AI Aristocracy
Commenters were struck by the $100,000 API cost for these discoveries, prompting discussions about the accessibility of such powerful AI tools. Some expressed concern that this financial barrier could lead to a 'tech aristocracy,' where only well-funded entities can leverage frontier AI, while others speculated about the superior throughput Anthropic's internal teams must have to consume tokens at such a rate.
PQC Peril & Prudence
The weakening of HAWK, a post-quantum cryptography candidate, fueled skepticism about the sole reliance on new PQC schemes. Commenters advocated for a continued hybrid approach, combining PQC with established classical ciphers, arguing that pushing for PQC-only solutions feels 'creepy' given the nascent nature and evolving understanding of these new algorithms, especially with AI now probing their weaknesses.
Autonomous Agents & AI's Quirks
The multi-agent workflow and Claude's initial 'reluctance' to tackle the AES problem, requiring human prompting, fascinated the audience. Discussion revolved around whether this 'collaboration' was true emergent intelligence or merely the stochastic nature of various search threads. Some pondered if LLMs, trained on human-like data, might exhibit 'tropey' behaviors like characters in fiction, where one agent's dismissal leads another to a breakthrough.
Disclosure Debates & Deployment Dilemmas
The article's closing remarks on how to react if an AI discovers vulnerabilities in *currently deployed* cryptosystems sparked a debate. Commenters questioned if the process for AI-discovered flaws should differ from human-discovered ones, pondering who to inform first (governments, public, etc.) and the potential national security implications, contrasting the rapid advancements of AI with traditional vulnerability disclosure timelines.