Frontier LLMs Demonstrate Cryptanalysis Capability, Uncover New Attacks on Established Cryptographic Primitives
What Happened — Researchers released CryptanalysisBench, a 191‑task benchmark that measures large language models’ ability to discover cryptanalytic attacks. Five frontier models (Claude Opus 4.8, Sonnet 5, Mythos 5, GPT‑5.5, GLM‑5.2) broke 65‑86 % of Tier 1 schemes and produced novel attacks against the SpoC AEAD and KINDI CCA‑security proof. Anthropic’s own Mythos model uncovered new vulnerabilities in the Hawk protocol and a reduced‑round AES variant.
Why It Matters for Compliance & Audit Readiness
- SOC 2 CC6 (Cryptographic Controls) requires that encryption algorithms be vetted, documented, and continuously monitored for emerging weaknesses.
- AI‑driven cryptanalysis introduces a new, fast‑moving attack surface that must be reflected in your risk‑assessment and control‑evidence processes.
- Verisq’s Control Mapping capability can automatically map cryptographic controls to SOC 2 criteria and capture continuous evidence that your chosen primitives remain resilient against AI‑generated attacks.
Who Is Affected — Technology‑SaaS providers, cloud‑infrastructure operators, fintech platforms, and any organization that relies on standard cryptographic primitives for data protection.
Recommended Actions
- Inventory every cryptographic algorithm in use and map each to SOC 2 CC6 control requirements.
- Incorporate AI‑driven cryptanalysis (e.g., CryptanalysisBench) into your periodic security‑testing program.
- Capture and retain evidence of algorithm vetting, test results, and remediation decisions as part of your continuous‑compliance audit trail.
Source: Schneier on Security – Measuring LLMs’ Ability to Perform Cryptanalysis
Technical Notes — The benchmark covers block ciphers, hash functions, AEAD schemes, and other primitives drawn from NIST competitions. No CVE identifiers are involved; the risk stems from novel mathematical attacks discovered by LLMs, not from a disclosed software flaw.