# Anthropic cryptographic weaknesses with Claude source note — 2026-08-03

Purpose: weekly Managing Expectations AI Papers Library maintenance note. Selected item: Anthropic's July 28, 2026 Frontier Red Team post, `Discovering cryptographic weaknesses with Claude`, and its connected technical papers/benchmark.

## Bottom line

Substantive source found and selected for one new library/blog note. Anthropic published a primary Frontier Red Team post on July 28, 2026 saying Claude Mythos Preview helped discover improved attacks on two cryptographic research targets:

- HAWK, a post-quantum digital-signature candidate in NIST's additional-signatures process.
- A reduced-round version of AES-128, not the full production AES cipher.

Editorial framing: this is a meaningful AI capability/safety signal because it moves frontier-model evaluation into high-stakes mathematical security research. It should **not** be framed as `Claude broke internet encryption`: Anthropic explicitly says neither result has practical impact on today's production systems; HAWK is a candidate scheme and the AES result concerns a reduced 7-round version, not full 10-round AES-128.

## Sources checked this run

### Anthropic Research feed

- Page: https://www.anthropic.com/research
- Accessed: 2026-08-03
- Tooling: direct HTML retrieval via Python/urllib.
- Recent visible items included:
  - `Discovering cryptographic weaknesses with Claude` — Jul 28, 2026.
  - `Project Pilot: Can AI control a drone?` — Jul 24, 2026. Already captured locally in `research/ai/anthropic-project-pilot-drone-bench-source-note-2026-07-27.md` and `blog/articles/anthropic-drone-bench-physical-agent-safety.html`.
  - `How Canada uses Claude: Findings from the Anthropic Economic Index` — Jul 14, 2026.
  - `Claude's values across models and languages` — Jul 13, 2026.
  - `Claude plays robotics` — Jul 9, 2026.
  - `An off switch for dual-use knowledge in AI models` — Jul 8, 2026. Already captured locally.
  - `A global workspace in language models` — Jul 6, 2026.

Selection rationale: the July 28 cryptography post was newer than the existing July 27 Managing Expectations AI-library article and had no matching local article/source note.

### Anthropic article / official source

- Page: https://www.anthropic.com/research/discovering-cryptographic-weaknesses
- Accessed: 2026-08-03
- Page title: `Discovering cryptographic weaknesses with Claude \ Anthropic`
- Page headline: `Discovering cryptographic weaknesses with Claude`
- Page date: Jul 28, 2026
- Source category on page: Frontier Red Team
- Description metadata: `Anthropic researchers find weaknesses in cryptographic algorithms with Claude Mythos Preview`

Key facts verified from the page text:

- Anthropic says Claude Mythos Preview helped discover improved attacks against cryptographic algorithms themselves, not only implementation bugs in cryptographic libraries.
- Anthropic's summary states that the first attack significantly weakens HAWK, a post-quantum digital-signature scheme, and the second identifies a new way to attack round-reduced AES.
- Anthropic explicitly says neither result has practical impact on today's computer systems and that no production software has to change as a result.
- Anthropic says HAWK is a candidate signature scheme and is not deployed.
- Anthropic says the AES result is on a reduced version of AES and does not break the full cipher.
- Anthropic says the two main findings cost roughly $100,000 in API cost each to develop.
- Anthropic says the HAWK finding was developed semi-autonomously with one researcher over about 60 hours; the AES finding was discovered almost entirely autonomously in a scaffold where Claude posed hypotheses and ran experiments.
- Anthropic says it followed responsible disclosure procedures, consulted academics, shared advance copies with US government and industry partners, and coordinated disclosure of the HAWK attack with the HAWK authors/public NIST mailing list.
- Anthropic says it partnered with academics at ETH Zurich, Tel Aviv University, and TU Berlin on CryptanalysisBench, a benchmark for evaluating LLM cryptanalytic ability.

### HAWK paper / Anthropic PDF

- PDF: https://anthropic.com/document/hawk_key_recovery.pdf
- Accessed/downloaded for text check: 2026-08-03
- PDF size observed locally: 748,154 bytes
- Title extracted from first page: `HAWK-n Key Recovery Reduces to SVP in Dimension n/2 + 1`
- Authors extracted from first page: Zygimantas Straznickas and Stephen A. Weis, Anthropic
- Key abstract points checked:
  - The paper describes HAWK as a lattice signature scheme currently a third-round candidate in NIST's post-quantum signature competition.
  - It gives a deterministic polynomial-time reduction from HAWK-n key recovery to exact SVP oracle calls in dimension n/2 + 1.
  - It states gate-count-model reductions for HAWK-512 from 2^150 to 2^108 and HAWK-1024 from 2^288 to 2^182.
  - It says the construction does not transfer to Falcon.

### AES paper / Anthropic PDF

- PDF: https://anthropic.com/document/aes_mobius_bridge.pdf
- Accessed/downloaded for text check: 2026-08-03
- PDF size observed locally: 1,084,840 bytes
- Title extracted from first page: `Cryptanalysis of 7-Round AES via the Algebraic Structure of its S-box`
- Authors extracted from first page: Milad Nasr and Nicholas Carlini, Anthropic
- Key abstract/introduction points checked:
  - The paper is about 7-round AES in the single-key setting, not full 10-round AES-128.
  - It introduces an algebraic `Möbius bridge` to reduce one guessed key byte and lower attack runtime in a reduced-round setting.
  - It says the improvements were discovered autonomously by a large language model and that human researchers focused heavily on validation.

### CryptanalysisBench / arXiv

- API: https://export.arxiv.org/api/query?id_list=2607.18538
- Page: https://arxiv.org/abs/2607.18538
- Accessed: 2026-08-03
- Title: `CryptanalysisBench: Can LLMs do Cryptanalysis?`
- Authors: Lukas Fluri, Avital Shafran, Nicholas Carlini, Matthew Jagielski, Milad Nasr, Orr Dunkelman, Eyal Ronen, Florian Tramèr
- Submitted: 2026-07-20; last revised: 2026-07-29 (v2)
- Category: cs.CR
- DOI: https://doi.org/10.48550/arXiv.2607.18538
- Abstract points checked:
  - CryptanalysisBench includes 191 tasks across six families of cryptographic primitives, primarily from four NIST standardization competitions.
  - It has tiers for known practical breaks, schemes with no known practical break/full-strength and scaled-down variants, and a challenge set of production primitives at the frontier of cryptanalysis.
  - The abstract reports five frontier models breaking 65%–86% of Tier 1 schemes and finding some novel cryptanalysis.

### Other watch lanes checked

#### OpenAI official RSS

- Feed: https://openai.com/news/rss.xml
- Accessed: 2026-08-03
- Recent RSS items included:
  - `Ten advances in mathematics and theoretical computer science` — Aug 1, 2026.
  - `Building abundant intelligence` — Jul 31, 2026.
  - `Disrupting a Criminal Scam Operation` — Jul 31, 2026.
  - `Advancing the price-performance frontier with GPT-5.6` — Jul 30, 2026.
  - `How enabling two settings tripled our scores on the ARC-AGI-3 benchmark` — Jul 29, 2026.
- Direct OpenAI news page check returned HTTP 403 in this runtime, so OpenAI items were treated as official RSS leads only and not selected for a new article.

#### Google DeepMind / Google Research

- DeepMind blog: https://deepmind.google/blog/
- Google Research blog: https://research.google/blog/
- Accessed: 2026-08-03
- Google Research visible recent item selected as a watch lead but not this week's article: `Science One Framework: A verifiable autonomous research framework via Chain-of-Evidence` — July 30, 2026.
- Selection rationale: the Anthropic cryptography item was more directly in the frontier-lab safety/capability lane and had multiple primary technical source documents available.

#### Meta AI

- Page: https://ai.meta.com/blog/
- Accessed: 2026-08-03
- Recent visible items included Muse Spark / Muse image-video / assistive robotics / Brain2Qwerty items. No item was selected this run.

#### LawZero / Yoshua Bengio

- Newsroom: https://lawzero.org/en/newsroom
- Blog: https://lawzero.org/en/blog
- Accessed: 2026-08-03
- Recent item visible: `LawZero advances safe-by-design AI with support from NVIDIA`, which was already captured in `research/ai/yoshua-bengio-update-2026-08-02.md` and summarized on `ai.html`.

#### Roman Yampolskiy

- Official YouTube RSS: https://www.youtube.com/feeds/videos.xml?channel_id=UCPIq6Bb-1iLmqyksJjy4kLQ
- Accessed: 2026-08-03
- Latest official-feed entries remained the July 4, 2026 short and July 1, 2026 Roman Forum interview, consistent with `research/ai/roman-yampolskiy-update-2026-08-01.md`.

## Why this was selected

- It is a primary-source Frontier Red Team release from a frontier AI lab.
- It has two linked technical papers, a benchmark paper, and responsible-disclosure framing.
- It directly concerns a high-stakes domain where frontier model capability could have both defensive and dangerous implications.
- It requires careful public explanation because the headline can easily be exaggerated into the false claim that production internet encryption was broken.

## Article framing used

Title: `Claude and Cryptography: A Real Warning, Not a Broken-Internet Panic`

Evidence label: `frontier red-team cryptanalysis / AI security capability warning`

Core caution: Claude-assisted cryptanalysis is a serious capability signal for digital-security research, but the HAWK and reduced-round AES results do not currently affect production systems and should not be converted into panic or marketing hype.

## Files updated locally

- `blog/articles/anthropic-cryptographic-weaknesses-claude.html`
- `research/ai/anthropic-cryptographic-weaknesses-claude-source-note-2026-08-03.md`
- `ai-library.html`
- `ai.html`
- `research/ai/ai-paper-library-seed-2026-06-11.json`
- `research/ai/ai-paper-library-source-note-2026-06-11.md`
- `blog/index.html`
- `sitemap.xml`