Public story · 2026-07-22 · high

AI Models Break a Cipher, Expose a Proof Error

CryptanalysisBench had five frontier models break 65 to 86 percent of already-broken ciphers across 191 tasks.

Why now: The arXiv paper is part of the July 22 coverage of frontier-model capability claims, one of the few dated data points behind the AI-does-real-research argument.

Issue

2026-07-22

Confidence

high

Sources

Redaction

passed

Story

Five frontier models cracked 6 to 12 previously unbroken cryptographic schemes at full strength, per a benchmark called CryptanalysisBench. Two of those cracks were new to the field: a key-recovery vulnerability in SpoC AEAD and an error in KINDI's security proof. For anyone maintaining or reviewing a custom cipher, that's the review process failing at its one job.

The benchmark ran 191 tasks across six families of cryptographic primitives, posted to arXiv as 2607.18538. It broke 65 to 86 percent of ciphers already known to be broken. Scaled-down versions of the hardest problems fell 24 to 61 times, tested across Claude Opus 4.8, Sonnet 5, Mythos 5, GPT 5.5, and GLM 5.2.

Yes, but reproducing a textbook attack is pattern matching, and that's most of the 65 to 86 percent.

The arXiv post doesn't say whether SpoC AEAD's or KINDI's maintainers have been notified, or whether either flaw holds up outside the benchmark's controlled conditions.

Source trail

arXiv 2607.18538

Entities

Claude Opus CryptanalysisBench Frontier GLM GPT KINDI LLM Models Mythos Sonnet SpoC AEAD

Claim evidence

AI Models Break a Cipher, Expose a Proof Error
Source

Provenance

Canonical issue: 2026-07-22
AI generated: yes
Story unit: 2026-07-22-frontier-models-break-6586-of-known-broken-ciphers-and-found-genuinely-novel-attacks
Labels: source-backed, canonical briefing excerpt

Fetching from the wire…

AI Models Break a Cipher, Expose a Proof Error

Story

Related stories

GLM 5.2 and the coming AI margin collapse

An Open-Source Model Just Beat GPT-5.4 and Opus 4.6 on the Hardest Coding Benchmark

SWE-bench Is Contaminated. The Primary Coding Agent Benchmark Is Broken.

OpenCode: Open-Source AI Coding Agent Hits 802 HN Points, 95K+ Stars — The Real Claude Code Rival Has Arrived

Poisoned Memory Files Reinfect Claude Code, Codex Sessions

GPT-5.5 Is Here. It's Fast, It's Expensive, and It Changes How You Prompt.

Commerce pulled two Anthropic models offline, then restored them

Frontier capability and general availability are now decoupled by policy.

Source trail

Entities

Claim evidence

Provenance

AI Models Break a Cipher, Expose a Proof Error

Story

Related stories

GLM 5.2 and the coming AI margin collapse

An Open-Source Model Just Beat GPT-5.4 and Opus 4.6 on the Hardest Coding Benchmark

SWE-bench Is Contaminated. The Primary Coding Agent Benchmark Is Broken.

OpenCode: Open-Source AI Coding Agent Hits 802 HN Points, 95K+ Stars — The Real Claude Code Rival Has Arrived

Poisoned Memory Files Reinfect Claude Code, Codex Sessions

GPT-5.5 Is Here. It's Fast, It's Expensive, and It Changes How You Prompt.

Commerce pulled two Anthropic models offline, then restored them

Frontier capability and general availability are now decoupled by policy.

Source trail

Entities

Claim evidence

Provenance