Anthropic's Claude Mythos Preview has discovered mathematical weaknesses in cryptographic algorithms that human experts missed for years, including a flaw in a leading post-quantum signature scheme under consideration by the U.S. government. The findings highlight both the remarkable capability of advanced AI systems and an emerging bottleneck: validating AI-generated research faster than humans can keep up.

What Did Claude Mythos Actually Discover?

Anthropic published research showing that Claude Mythos Preview uncovered two significant cryptographic vulnerabilities. The first involved HAWK, a post-quantum digital signature scheme that is a third-round candidate in the National Institute of Standards and Technology's (NIST) call for additional post-quantum signatures. HAWK had survived two rounds of expert human review over two years, yet Mythos found a hidden mathematical symmetry, called an automorphism, in the lattice structure HAWK uses. Prior theoretical work had proven such a symmetry would enable an attack, but no one had actually found one until now.

The attack discovered by Mythos cuts HAWK's effective key strength in half, meaning the scheme would need double the key size to reach its claimed security level. While the attack is not fast enough to break HAWK outright, it undermines one of the scheme's main advantages: compact key sizes. A single Anthropic researcher with a theoretical computer science background, not a lattice-cryptography expert, worked with Mythos over 60 hours at a cost of roughly $100,000 in API tokens.

The second discovery involved a reduced seven-round variant of AES-128, the encryption standard used worldwide. Mythos discovered a technique called a "Mobius Bridge" that improves the best known meet-in-the-middle attack by 200 to 800 times. This attack works only under a chosen-plaintext threat model and poses no practical risk to real systems using full AES, which employs 10 rounds. However, the discovery process was remarkable: Mythos worked almost entirely autonomously over three days, producing roughly one billion output tokens, with only three substantive human prompts.

Why Is the Human Validation Bottleneck the Real Story?

The most striking finding may not be what Mythos discovered, but what happened next. After Mythos produced the core insight for the AES attack in three days, two human researchers spent several hundred hours validating the result. Anthropic acknowledged they are "reaching the limits of our own knowledge" on verifying Claude's cryptographic results. This mirrors a broader pattern: as AI systems generate research output faster, the bottleneck shifts from generating ideas to validating them.

Anthropic

The company also announced CryptanalysisBench, a benchmark built with ETH Zurich, Tel Aviv University, and University of Haifa that packages cryptographic ciphers for evaluating large language model (LLM) cryptanalytic capabilities. Additional results showed practical attacks on 13 of 24 rounds of LEA cipher running in under an hour on a desktop computer, plus smaller improvements on Serpent-128, Salsa20, Poseidon, and SHA-1.

What Does the Developer Community Think?

The Hacker News discussion revealed sharp disagreement about the significance of these findings. Top commenters noted important context: the AES attack works only on a reduced seven-round variant, not the full cipher, and uses a threat model that is no longer realistic in modern encryption modes like AES-GCM, which use counter mode where attackers cannot influence what gets encrypted. The HAWK attack, however, drew more serious attention because HAWK is a candidate for government standardization and has not been widely deployed.

The $100,000 cost sparked debate about access inequality. One commenter calculated that $100,000 in tokens in a week represents an impressive feat even with massive parallelization, and speculated that Anthropic's internal researchers have access to far higher token throughput than public endpoints. This raised concerns about a "tech aristocracy" emerging where only well-funded organizations can afford to run such experiments. Other commenters questioned whether Chinese AI models could reproduce these results at lower cost.

How Did Messy Prompts Lead to Breakthrough Discoveries?

One of the most counterintuitive findings was how Anthropic's researchers prompted Mythos. The prompts contained typos and grammatical errors, such as "no again the goal is that we have highly inteligent model as good top researcher, we want to find new attacks." This sparked a meta-discussion in the developer community about whether elaborate prompt engineering and complex system instructions are overrated. Several commenters argued that giving an LLM a bash tool and a simple prompt often outperforms complex setups with skills files and detailed instructions.

The discovery process also revealed interesting multi-agent dynamics. Two worker agents operated in parallel; one prematurely rejected a key idea while the other found a way to exploit it, and they eventually converged on the solution. This raised questions about whether the collaboration was genuine or simply stochastic search, and whether replaying the experiment would consistently find the same path.

Steps to Understand the Implications for Cryptography and AI Safety

• Immediate Impact Assessment: Anthropic emphasized that none of these results have practical impact on production systems today, since HAWK is not yet deployed and the AES attack targets a reduced variant under unrealistic threat conditions.
• Future Preparedness: The company warned that as model capabilities grow, organizations should prepare for a future where LLMs find real-world cryptographic vulnerabilities faster than humans can validate them, creating a validation bottleneck.
• Benchmark Development: CryptanalysisBench provides a standardized way to evaluate how well different LLMs can discover cryptographic weaknesses, enabling researchers to track progress and compare models over time.
• Research Validation Pipelines: The findings suggest that AI-assisted verification systems may be necessary before AI-generated research output accelerates further, since human validation alone cannot keep pace.

What Are the Broader Implications?

The research raises a chilling long-term concern: as AI systems transmute computational tokens into research effort, the world may split into two categories of problems. Some problems will yield to AI attack, making human effort entirely unnecessary. Others will harden to the point where human effort feels increasingly less worthwhile, because "even AI couldn't solve it." This could reshape how cryptographers and security researchers approach their work.

Anthropic's findings also underscore a shift in how advanced AI systems are being used. Rather than serving as tools that humans direct with carefully engineered prompts, systems like Mythos are operating with significant autonomy, generating billions of tokens over days with minimal human intervention. The messy, unpolished prompts that worked best suggest that intent and context matter more than precision, a lesson that could reshape how developers build AI-powered research systems.

The $100,000 cost and the validation bottleneck together point to an emerging reality: cutting-edge AI research may increasingly require both massive computational resources and equally massive human effort to validate results. This could concentrate cryptographic research capability among well-funded organizations, raising questions about who gets to shape the future of cryptographic standards and security.