Overview
A new academic benchmark, CryptanalysisBench, published in July 2026 demonstrates that state-of-the-art large language models are no longer merely reciting known cryptanalytic results — they are beginning to generate novel ones. Five frontier models, including Anthropic’s Claude Opus 4.8 and Mythos 5, OpenAI’s GPT-5.5, and the open-weights GLM-5.2, were evaluated across 191 tasks spanning six families of cryptographic primitives drawn from four NIST standardisation competitions. The results represent a meaningful inflection point: AI is beginning to operate at the frontier of one of the most technically demanding domains in security.
Technical Analysis
CryptanalysisBench is structured across three tiers of difficulty:
- Tier 1: Primitives with known practical breaks. Frontier models achieved a 65–86% break rate, demonstrating strong recall and application of established cryptanalytic knowledge.
- Tier 2: Primitives with no known practical break, tested at full strength and in scaled-down variants. Models broke 6–12 full-strength schemes and 24–61 scaled-down variants — a result that goes well beyond reproducing literature.
- Tier 3: A challenge set of production primitives at the frontier of human cryptanalysis.
Most significantly, Anthropic’s Mythos Preview model produced genuinely novel results: a key-recovery attack exploiting a design flaw in the SpoC AEAD scheme, identification of an error in KINDI’s published CCA-security proof, and new vulnerabilities in the Hawk signature scheme and reduced-round AES. These are described by the authors as previously unknown findings — the first time an AI system has demonstrably advanced the published state of the art in cryptanalysis.
The automatic verifiability of cryptanalytic attacks (a correct attack either works or it doesn’t) makes this domain uniquely well-suited to LLM-driven research loops, potentially enabling rapid, automated discovery pipelines.
Framework Mapping
- AML.T0040 / AML.T0044: The models are being used in full-access inference settings to reason over cryptographic specifications, analogous to adversarial use of ML model capabilities against security-critical infrastructure.
- AML.T0047: The capability is delivered via frontier model APIs and products, meaning the attack surface is accessible to a wide range of threat actors.
- LLM08 – Excessive Agency: As LLMs gain the ability to autonomously discover and verify novel attacks, the risk of autonomous cryptanalytic pipelines operating without human oversight increases materially.
- LLM09 – Overreliance: Defenders and standardisation bodies may be overreliant on the assumption that only human experts can find novel weaknesses in cryptographic schemes.
Impact Assessment
The immediate impact is concentrated on cryptographic schemes in active standardisation or pre-deployment review — particularly NIST competition candidates. Legacy primitives with known structural weaknesses are at elevated risk of accelerated exploitation if adversaries deploy similar AI-assisted cryptanalysis pipelines. Nation-state actors and well-resourced cybercriminal groups are the most plausible early adopters of offensive variants of this capability.
Mitigation & Recommendations
- Adopt AI-assisted pre-deployment review: Use tools like CryptanalysisBench to stress-test candidate schemes before production deployment.
- Monitor benchmark publications: Track CryptanalysisBench updates for any primitives currently in use or under consideration.
- Accelerate expert review of AI-flagged schemes: Treat LLM-generated cryptanalytic findings as credible leads requiring immediate expert validation, not speculative outputs.
- Update threat models: Cryptographic threat models should now account for AI-assisted adversaries capable of finding novel structural weaknesses.