Anthropic, a frontier artificial intelligence company, announced on July 28, 2026, that its Claude Mythos Preview AI model has uncovered significant cryptanalytic attacks, including potential weaknesses in Anthropic AI encryption.
These discoveries target both weakened versions of the widely used Advanced Encryption Standard (AES) and HAWK, a promising post-quantum digital signature scheme currently under review by the National Institute of Standards and Technology (NIST).
Anthropic AI encryption vulnerabilities discovered
The findings have ignited fresh discussions within the cryptocurrency community about the long-term cryptographic security of assets like Bitcoin, though Anthropic stresses there’s no immediate threat to existing production systems.
This development adds an unexpected layer to the ongoing race for quantum-resistant cryptography, suggesting that artificial intelligence could potentially challenge future encryption methods even before quantum computers become a pervasive threat. It forces developers across various sectors, including blockchain, to consider a dual-pronged future for cryptographic security.
The Claude Mythos Preview model, working in concert with human researchers, demonstrated an impressive capability for original cryptographic research. It specifically identified previously unknown attacks against a weakened version of AES, reducing the time to break the algorithm by a factor of 200 to 1,000. This particular attack targeted AES with seven of its usual ten scrambling rounds.
More notably for the future of encryption, the AI also devised an improved attack against HAWK, a lattice-based post-quantum cryptographic signature scheme. This attack exposed a nontrivial automorphism within HAWK’s lattice structure, effectively halving its security. For the smallest HAWK 256 configuration, the estimated cost of recovering a key plummeted from approximately 264 operations to just 238 operations.
Anthropic’s researchers detailed the process, noting that Claude developed the core HAWK attack in about 60 hours using a multi-agent research system. The released implementation of this attack has an expected end-to-end runtime of about three hours and 42 minutes on a 96-core server.
The cost for the model run alone was approximately $100,000 in API usage, and human researchers spent several hundred hours validating the AI’s findings.
In a blog post detailing the work, Anthropic called the discoveries a “substantial” research advancement. They emphasized that the flaws do not affect software currently in use.
Still, the company posed a pertinent question: “As we develop increasingly powerful cryptanalytic results, it would be prudent to consider how researchers should react if a language model were to discover vulnerabilities in cryptosystems where attacks do have an immediate real-world impact.”
Bitcoin’s cryptographic migration faces new considerations
Bitcoin’s underlying security hinges on the Elliptic Curve Digital Signature Algorithm (ECDSA) for transaction authorization. Cryptographers have long recognized that sufficiently powerful quantum computers could eventually compromise ECDSA, leading to efforts to transition Bitcoin and other blockchain networks to post-quantum signature schemes. Lattice-based cryptography has emerged as a leading candidate for this transition.
The Anthropic findings introduce an intriguing twist to this narrative. Coin Center Executive Director Peter Van Valkenburgh lightheartedly mused about the “funniest timeline,” suggesting that Bitcoin could upgrade to a lattice-based post-quantum scheme only for advanced large language models to break its mathematics before quantum computers ever pose a threat. This highlights a new, parallel vector of vulnerability that blockchain developers must now consider.
Charles Guillemet, a Bitcoin Core Developer, echoed these sentiments by noting that Anthropic’s research points to AI-assisted cryptanalysis significantly reducing the complexity of a known private key recovery attack against the HAWK signature scheme. He advocated for “crypto-agility,” the ability to swiftly adapt and swap cryptographic algorithms, as an essential defense mechanism against such rapidly evolving threats.
NIST’s post-quantum evaluation process validated
While Anthropic’s revelations might seem alarming, some experts view them as a testament to the effectiveness of current cryptographic review processes. The HAWK scheme, for instance, was still under evaluation by NIST, having only moved into the third round of its post-quantum signature competition in May 2026. Anthropic had notified HAWK’s designers of the discovered weakness in June 2026, well before the public announcement.
Ellen Boehm, Senior Vice President of Strategy and AI Innovation at Keyfactor, stated that research like Anthropic’s proves the NIST Post-Quantum Cryptography (PQC) evaluation process is functioning as intended. She also underlined the heightened importance for organizations to maintain clear visibility into their cryptographic deployments across their enterprise systems and processes.
The fact that the AI attacks targeted deliberately weakened cryptographic systems, rather than full production-grade implementations, provides a crucial buffer. Anthropic explicitly confirmed that modern AES remains secure, underscoring that the research’s intent was to advance understanding and identify potential vulnerabilities in controlled environments.
Accelerating quantum and AI threats to encryption
The long-standing concern about quantum computing breaking existing encryption is intensifying. Google plans to switch to post-quantum cryptography by 2029, while US agencies face a 2031 quantum deadline. Compounding this, Google Quantum AI researchers demonstrated on March 31, 2026, that breaking elliptic curve cryptography, the foundation of Bitcoin’s security, might require fewer than 500,000 physical qubits—a 20-fold reduction from previous estimates.
Justin Drake of the Ethereum Foundation estimates there’s at least a 10% chance that a quantum computer could recover a private key from an exposed public key by 2032. This timeline poses a substantial risk for Bitcoin, with projections indicating that quantum computing could compromise its encryption by 2029.
This potential vulnerability could expose up to 35% of Bitcoin’s circulating supply, particularly those held in older wallets with exploitable cryptographic data.
The convergence of advanced AI capable of cryptanalysis with the accelerating progress in quantum computing presents a complex, multi-faceted challenge. It implies that future cryptographic systems won’t just need to withstand the brute force of quantum algorithms but also the subtle, emergent intellectual capabilities of artificial intelligence. This dual threat necessitates a proactive and adaptable approach from developers and security professionals across the digital landscape.
Preparing for a cryptographically uncertain future
While the immediate threat to Bitcoin and the broader internet remains low, Anthropic’s findings serve as a potent warning. The rapid advancement of AI’s capabilities in mathematical discovery means that developers of blockchain networks and other critical digital infrastructure must begin evaluating post-quantum security not just against theoretical quantum attacks, but also against increasingly sophisticated AI cryptanalysis.
This isn’t about panic, but rather preparedness. The cryptographic community, alongside AI researchers, will need to collaborate closely to ensure new encryption standards are robust against both known and emerging computational threats. It’s a continuous arms race where the defense must always anticipate the next generation of offense, whether it comes from qubits or neural networks.