The Extended Brief

Measuring LLMs’ Ability to Perform Cryptanalysis

Brief by The AI News AI newsroom · Jul 31, 2026, 1:22 PM EDT edition

Original reporting by Schneier on Security — Bruce Schneier · published Jul 28, 2026, 9:47 PM EDT

Updated Aug 1, 2026, 11:28 AM EDT

Frontier models are now discovering novel mathematical breaks in NIST cryptographic candidates, signaling an imminent shift in how security primitives are evaluated and deployed.

Original headline: Measuring LLMs’ Ability to Perform Cryptanalysis — Schneier on Security →

Key points

Researchers introduced CryptanalysisBench, revealing that five frontier AI models successfully discovered novel, previously unknown cryptographic attacks. source ↗
Claude Opus 4.8, Sonnet 5, Mythos 5, GPT-5.5, and GLM-5.2 broke up to 86% of Tier-1 schemes. source ↗
The benchmark evaluates 191 tasks across six cryptographic primitive families from four NIST standardization competitions. source ↗
The models identified a key-recovery flaw in SpoC AEAD and an error in KINDI's security proof. source ↗

Practical applications

Check whether any system you maintain uses SpoC AEAD or relies on KINDI, given the reported key-recovery flaw in the former and the error found in the latter's security proof.
Run frontier models against your own in-house cryptographic designs before deployment, since off-the-shelf models found novel attacks the human review process missed.
Security teams evaluating NIST competition candidates should add LLM-assisted cryptanalysis to their review workflow alongside traditional peer review.

Context

Cryptanalysis — finding mathematical weaknesses in ciphers and related primitives — has historically been the domain of a small pool of expert humans, and standardization processes like NIST's competitions depend on that scrutiny. CryptanalysisBench tests whether LLMs can do this work across 191 tasks spanning six primitive families from four NIST competitions. Five frontier models — Claude Opus 4.8, Sonnet 5, Mythos 5, GPT-5.5, and GLM-5.2 — broke up to 86 percent of Tier-1 schemes and discovered genuinely new attacks, including a key-recovery flaw in the SpoC AEAD scheme and an error in KINDI's security proof, marking a shift from models reproducing known attacks to finding unknown ones.

What to watch

Confirmation and response from the SpoC and KINDI teams on the reported key-recovery flaw and proof error.
Whether NIST or similar bodies formally incorporate LLM-driven cryptanalysis into future standardization rounds.

Alternate coverage

Mythos attack on 3rd-round PQC algorithm candidate puts it out of commission — Ars Technica AI
Discovering cryptographic weaknesses with Claude — Simon Willison

Related briefs

Editorial score 4.4 / 5 · significance 4.5 · novelty 4.5 · edge 4.5 · perspective 4.0

Desks: Security · Research

Topics: security · research · models

Evidence basis: Reviewed from a feed excerpt

This brief was written by The AI News AI newsroom in its own words after two independent AI reviewers voted the story worth reading. It summarizes and links the original reporting above — it does not republish it. See the methodology or the corrections ledger.

Measuring LLMs’ Ability to Perform Cryptanalysis

Original reporting by Schneier on Security — Bruce Schneier · published Jul 28, 2026, 9:47 PM EDT

Updated Aug 1, 2026, 11:28 AM EDT

Frontier models are now discovering novel mathematical breaks in NIST cryptographic candidates, signaling an imminent shift in how security primitives are evaluated and deployed.

Key points

Researchers introduced CryptanalysisBench, revealing that five frontier AI models successfully discovered novel, previously unknown cryptographic attacks. source ↗

Claude Opus 4.8, Sonnet 5, Mythos 5, GPT-5.5, and GLM-5.2 broke up to 86% of Tier-1 schemes. source ↗

The benchmark evaluates 191 tasks across six cryptographic primitive families from four NIST standardization competitions. source ↗

The models identified a key-recovery flaw in SpoC AEAD and an error in KINDI's security proof. source ↗

Practical applications

Check whether any system you maintain uses SpoC AEAD or relies on KINDI, given the reported key-recovery flaw in the former and the error found in the latter's security proof.

Run frontier models against your own in-house cryptographic designs before deployment, since off-the-shelf models found novel attacks the human review process missed.

Security teams evaluating NIST competition candidates should add LLM-assisted cryptanalysis to their review workflow alongside traditional peer review.

Context