Key Takeaways
- Claude Mythos Preview identified a lattice symmetry in HAWK that cuts key-recovery effort from 2^64 to 2^38 operations
- The AI developed a new technique reducing seven-round AES attack complexity by 200-800x
- Neither finding affects production systems; full AES-128 and deployed encryption remain secure
Anthropic claims its Claude Mythos Preview model has discovered previously unknown mathematical weaknesses in two cryptographic algorithms. The company's Frontier Red Team detailed how the AI identified a structural flaw in HAWK, a post-quantum signature candidate, and developed a faster attack on reduced-round AES. Neither finding compromises systems currently deployed in the real world.
This marks one of the first public demonstrations of a frontier AI contributing novel cryptographic research. The results raise practical questions for fintech security teams: how soon will AI-assisted cryptanalysis become standard, and what does that mean for algorithm selection today?
What did Claude find in HAWK?
HAWK is a digital signature scheme designed to resist attacks from both classical and quantum computers. It is currently advancing through NIST's post-quantum cryptography standardization process, having survived two rounds of expert review spanning roughly two years.
Claude Mythos Preview identified a nontrivial symmetry in the underlying lattice mathematics that human reviewers had missed. This symmetry enables a significantly more efficient key-recovery attack. For HAWK-256, the smallest parameter set, the computational effort dropped from approximately 2^64 operations to roughly 2^38 operations.
That 26-bit difference roughly halves the effective key strength. Correcting the vulnerability would require larger keys, which would erode HAWK's performance advantage over competing post-quantum schemes. Since HAWK has not been deployed in any production environment, the finding creates no immediate operational risk. It does, however, complicate HAWK's standardization path.
The discovery took approximately 60 hours of semi-autonomous work, guided by a single researcher. Anthropic emphasizes that Claude performed the mathematical exploration largely on its own, with human scaffolding limited to tool access and literature retrieval.
How does the AES attack work?
AES-128, standardized by NIST in 2001, uses ten rounds of substitution and permutation. Cryptographers routinely study reduced-round variants to probe for weaknesses. Claude Mythos Preview targeted a seven-round version and developed a new technique the company calls the Möbius Bridge.
The Möbius Bridge eliminates one guessing step from prior meet-in-the-middle attacks. Combined with other optimizations the model generated, the new approach runs between 200 and 800 times faster than previous best results on seven-round AES, depending on the measurement criteria.
Full-strength AES-128 remains unaffected. The improved attack is also theoretical in a practical sense: it requires an impractically large volume of plaintext-ciphertext pairs. No deployed system faces new risk from this work.
Anthropic notes preliminary results on reduced-round variants of the Lightweight Encryption Algorithm (LEA) and Serpent, though these are still under investigation.
Explores enterprise trust concerns when deploying frontier AI systems
What did the work cost?
Each research effort incurred roughly $100,000 in API usage, according to Anthropic. The AES work required extensive human verification afterward, meaning the total cost including researcher time is likely considerably higher.
For context, academic cryptanalysis research rarely has access to that level of compute budget. If AI-assisted cryptanalysis scales, well-funded adversaries could probe encryption standards far more quickly than academic review processes can keep pace.
| Metric | HAWK Attack | Seven-Round AES Attack |
|---|---|---|
| Algorithm status | NIST candidate, not deployed | Reduced variant of deployed standard |
| Previous best effort | 2^64 operations | Baseline meet-in-the-middle |
| New effort | 2^38 operations | 200-800x faster |
| Real-world impact | None (undeployed) | None (full AES unaffected) |
| Human involvement | ~60 hours guidance | Extensive post-hoc verification |
| Approximate cost | $100,000 API | $100,000+ API plus verification |
Why this matters for fintech security teams
The findings themselves pose no operational threat. What they signal is more significant: frontier AI systems can now contribute original cryptographic research, identifying subtle mathematical structures that evade prolonged human analysis.
Fintech teams typically rely on NIST-approved algorithms and assume that multi-year standardization processes catch critical flaws. That assumption still holds for deployed standards like AES and RSA. But the HAWK result shows that even advanced candidates can harbor exploitable structure.
Post-quantum migration is already on the roadmap for most financial institutions. NIST finalized its first post-quantum standards in 2024, recommending ML-KEM and ML-DSA. HAWK was a signature alternative still in review. Teams planning migration should note that AI-assisted cryptanalysis may accelerate the discovery of weaknesses in newer algorithms, potentially shortening their usable lifespan.
Tracks deeptech investment trends including security and crypto startups
What Anthropic is not claiming
Anthropic frames the work as constructive research, not a threat demonstration. The company explicitly states that neither finding compromises any system in widespread use. The goal, according to the Frontier Red Team, is to strengthen future standards and accelerate defensive research.
The company plans additional disclosure of complete technical details after further review. That staged release is standard practice in responsible vulnerability disclosure, though it also gives Anthropic time to refine its claims before peer scrutiny.
Skeptics may note that Anthropic has commercial incentives to publicize frontier capabilities. The HAWK finding will need independent verification, particularly given that the lattice community has studied the scheme intensively. If the symmetry is real and significant, other researchers should be able to replicate the result.
Implications for cryptographic agility
Cryptographic agility, the ability to swap algorithms without re-architecting systems, has been a compliance checkbox for years. AI-assisted cryptanalysis makes it operationally urgent.
If novel attacks can emerge in weeks rather than years, organizations need infrastructure that can deprecate a compromised algorithm quickly. That means abstracting cryptographic primitives behind clean APIs, maintaining algorithm inventories, and monitoring NIST and academic channels for deprecation signals.
For fintech firms, where transaction authentication and data encryption are table stakes, cryptographic agility is no longer optional. The HAWK episode illustrates that even pre-deployment algorithms can become unsuitable faster than expected.
Logicity's Take
Anthropic's announcement is part research milestone, part marketing. The HAWK result is genuinely interesting, but it affects a candidate that was never deployed. The AES result is incremental against a toy variant. The real story is that AI-assisted cryptanalysis is now plausible, which shifts the economics of attack research. Fintech teams should treat this as a planning signal, not an emergency: accelerate cryptographic agility roadmaps, prefer NIST-finalized standards (ML-KEM, ML-DSA) over candidates still in review, and expect that AI-discovered vulnerabilities will become a recurring disclosure category.
Relevant to teams evaluating AI agent performance in production systems
Frequently Asked Questions
Does this affect encryption used in banking and payments today?
No. Neither finding compromises deployed systems. Full-strength AES-128/256 and current TLS implementations remain secure. HAWK was never in production.
Should we stop using AES?
No. The attack targets a reduced-round variant (seven rounds instead of ten). It has no bearing on AES as deployed.
What is HAWK and why does this matter?
HAWK is a post-quantum digital signature scheme in NIST's standardization pipeline. The vulnerability may affect its path to approval and require larger key sizes, reducing its performance edge.
How reliable are Anthropic's claims?
The company has not yet released full technical details. Independent verification by cryptographers will determine whether the HAWK symmetry is as exploitable as claimed.
What should security teams do now?
Prioritize cryptographic agility. Prefer finalized NIST standards (ML-KEM, ML-DSA) over candidates still in review. Monitor for AI-discovered vulnerabilities as a new disclosure category.
Need Help Implementing This?
Logicity can connect you with vetted security consultants specializing in post-quantum migration and cryptographic agility assessments. Reach out via our consulting request form.
Source: Crowdfund Insider
Huma Shazia
Senior AI & Tech Writer
Produced with AI assistance and reviewed by the Logicity editorial team. Learn more in our Editorial Policy.
