Breaking Lab srl
AI Enters the Cryptanalysis Arena
For decades, breaking encryption has been the exclusive domain of specialist mathematicians armed with deep domain knowledge and years of painstaking research. That picture is shifting. Anthropic has revealed that its Claude Mythos Preview model independently derived a full end-to-end key-recovery attack against HAWK-256, a post-quantum digital signature scheme built on hard lattice problems, and separately identified a substantially accelerated attack path against a reduced-round version of the AES-128 block cipher. The results, published alongside a working implementation, mark one of the most concrete demonstrations yet of large language models contributing meaningfully to offensive security research.
What HAWK-256 Is — and Why Breaking It Matters
HAWK is a lattice-based signature scheme designed specifically for the post-quantum era, intended to remain secure even against adversaries equipped with large-scale quantum computers. Post-quantum cryptography has become urgent as NIST finalises its standards and governments worldwide begin migrating critical infrastructure. HAWK-256 is among the candidates being closely evaluated. Claude’s attack exploits a previously overlooked symmetry property in the underlying lattice structure — a subtle mathematical relationship that human researchers had apparently not thought to weaponise. Anthropic has released the implementation, which achieves a full key-recovery run in approximately three hours and 42 minutes on a 96-core server. That is well within reach of a well-resourced attacker.
A 200x Speed Boost Against Seven-Round AES
The AES result is no less striking. The full AES-128 cipher runs ten rounds of substitution and permutation; reducing it to seven rounds is a standard research technique used to probe structural weaknesses without attacking production deployments directly. Claude identified an attack path that cuts the computational effort by a factor of between 200 and 800 compared to previous best-known methods. While seven-round AES is not deployed in the real world, improvements of this magnitude on reduced-round variants are the precursor to advances that eventually threaten full-strength ciphers. Cryptographers take these results seriously as leading indicators.
How the Model Did It
Anthropic has not published a full technical account of Claude’s methodology, but the company indicates the model was given access to relevant academic literature and asked to reason about the cryptographic constructions. The HAWK attack, in particular, required identifying and exploiting a structural symmetry that had gone unnoticed in previous analysis. This is precisely the kind of non-obvious pattern recognition where large language models, trained on vast corpora of mathematical and computer science text, might surface connections a human researcher could overlook simply through the sheer breadth of associations the model can draw simultaneously.
Responsible Disclosure and the Publication Decision
Anthropic notified the HAWK development team and relevant parties before publishing its findings, consistent with standard responsible disclosure norms. The company has also released its implementation openly, which it frames as enabling the cryptographic community to validate the results and accelerate any necessary revisions to HAWK before wider deployment. That decision will not be universally popular: releasing a working key-recovery tool lowers the barrier for malicious actors, even if the academic consensus holds that open publication ultimately benefits defenders more than attackers over the long run.
Editorial Analysis: A Threshold Moment for AI-Assisted Offence
These findings deserve attention beyond the cryptography community. For years, the debate around AI in security has focused on defenders — threat detection, log analysis, vulnerability triage. Claude’s cryptanalytic results shift the conversation. If a commercially available language model can independently identify novel attacks on post-quantum candidates and substantially accelerate known attacks on a global encryption standard, the assumption that cutting-edge cryptanalysis requires rare human expertise needs revisiting. This is not yet a crisis: HAWK-256 is not in widespread production use, and seven-round AES is not AES. But it is a preview. The pace at which AI capabilities are advancing suggests that the gap between “reduced-round research target” and “production cipher” may close faster than the cryptographic community’s traditionally slow, consensus-driven revision cycles can accommodate. The post-quantum transition just became more urgent.
