The cryptographic foundations that governments are building for a post-quantum world may be less solid than anyone assumed. Anthropic's restricted artificial intelligence model, Claude Mythos, has done something that teams of human cryptographers spent years attempting and failing to do: it found a genuinely new attack on a post-quantum signature scheme that was actively on the path toward U.S. federal standardization. The implications ripple far beyond an academic curiosity — they cut straight to the heart of how the global financial and communications infrastructure intends to survive the coming era of quantum computing.
Post-quantum cryptography, or PQC, refers to a class of cryptographic algorithms designed to resist attacks from quantum computers, which are expected to eventually render many of today's encryption standards obsolete. The U.S. National Institute of Standards and Technology has been running a years-long standardization process to identify and enshrine the algorithms that will protect federal systems, financial networks, and critical infrastructure going forward. Getting onto that standardization track is not a casual achievement — it means a scheme has already survived intensive scrutiny from the global cryptographic research community. Or so everyone thought.
Claude Mythos is what Anthropic describes as a "locked model," meaning it operates under access restrictions that distinguish it from publicly available versions of Claude. The fact that this discovery came from a controlled, gated system rather than an open research deployment raises its own questions about what such models are capable of when pointed at hard, formally structured problems. Cryptography — with its precise mathematical rules and clearly defined success conditions — is exactly the kind of domain where AI systems can operate with unusual rigor. Claude Mythos appears to have exploited that structural clarity to devastating effect.
What makes the finding particularly striking is not merely that an AI solved a hard problem, but that it solved one humans had explicitly tried and failed to crack over multiple years. The scheme in question is a post-quantum signature scheme, a category of algorithm used to verify authenticity and integrity of digital communications and transactions. Signature schemes are foundational to everything from secure browsing and software updates to blockchain transaction validation and cryptocurrency wallet security. A viable attack on a scheme of this type — especially one being groomed for federal adoption — is not a theoretical concern. It is a practical threat to deployed or soon-to-be-deployed infrastructure.
For the digital assets industry specifically, this development carries weight that cannot be overstated. Blockchain networks and cryptocurrency protocols have long operated on the assumption that their underlying cryptographic primitives are secure. Many projects, including those exploring long-term protocol design, have begun explicitly planning for post-quantum migration, treating NIST-standardized PQC candidates as the safe destination. If an AI model can find novel attacks on those very candidates before they are even finalized, the migration path itself becomes uncertain terrain. The industry cannot simply wait for government standardization and then copy-paste the approved algorithms — it must now treat even vetted candidates with a new layer of skepticism.
The discovery also reframes a broader debate about AI capability in high-stakes technical domains. There has been a persistent argument that large language models are fundamentally pattern-matching engines, impressive at surface-level tasks but incapable of genuine mathematical discovery. Claude Mythos finding a new cryptographic attack — not reproducing a known one, but constructing something novel — is a direct challenge to that characterization. Anthropic has been cautious about publicizing its most capable models, and the "locked" designation of Claude Mythos suggests the company is aware that raw capability at this level demands careful handling.
From a regulatory and standards perspective, this episode will likely force a reckoning. The NIST standardization process has always relied on the assumption that human cryptanalysts, given sufficient time and incentive, will surface critical flaws before an algorithm is baked into federal systems. Claude Mythos has now demonstrated that this human-centric model of cryptographic review may be insufficient. If AI systems can find attacks that years of expert human analysis missed, then standards bodies will need to either integrate AI-assisted cryptanalysis into their evaluation pipelines or accept that some vulnerabilities may only surface after standardization — a far more dangerous scenario.
The timing is uncomfortable. Quantum computing hardware is advancing, regulatory frameworks for PQC are in their final stages, and financial institutions are beginning to make concrete infrastructure decisions based on the assumption that the NIST-approved candidate pool is trustworthy. Claude Mythos has injected genuine uncertainty into that assumption at exactly the moment when the industry could least afford it. Whether this leads to a productive acceleration of AI-assisted cryptographic review or to a paralysis of confidence in the standardization process depends almost entirely on how Anthropic, NIST, and the broader security community respond in the coming months. The crack in the foundation has been found. What gets built on top of it is now the question.
Written by the editorial team — independent journalism powered by Bitcoin News.