HN Debrief

Discovering Cryptographic Weaknesses with Claude

  • AI
  • Security
  • Research
  • Infrastructure

Anthropic’s post describes an AI-assisted cryptanalysis project where Claude, running inside a scaffold that let many agent sessions explore in parallel, helped produce two publishable results. One improved the best known chosen-plaintext attack on 7-round AES-128, which is a weakened version of the standard 10-round cipher and not a practical break of deployed AES. The other found a concrete automorphism in HAWK-256, a post-quantum signature candidate, turning a previously understood theoretical opening into a stronger attack that roughly halves effective key strength for that variant. Anthropic says each result took on the order of $100,000 in API cost plus substantial human effort to validate and package.

If you run security or core infrastructure, do not read this as “AES is broken.” Read it as a sign that frontier labs can now throw large-scale automated search at hard technical problems, which raises the value of early review, hybrid deployments, and faster response plans for immature standards.

Discussion mood

Interested but skeptical. Most people thought the results were real and nontrivial, but were annoyed by the PR framing, careful to stress that deployed AES is not broken, and uneasy about how much the work depended on expensive internal access and heavy human scaffolding.

Key insights

  1. 01

    HAWK and AES need plain-English framing

    The technical significance lands very differently once you strip away the headline language. HAWK-256 is a candidate post-quantum signature scheme, not something widely deployed today, and the AES result applies to a reduced-round version under stronger-than-normal assumptions. That changes the story from emergency breakage to useful but bounded cryptanalysis.

    Translate these claims for your team before they spread internally as “AI broke AES.” If you track cryptography risk, separate mature deployed primitives from candidate schemes and toyed-with reduced variants.

      Attribution:
    • _ache_ #1
    • reader9274 #1
    • baxtr #1
  2. 02

    The HAWK result looks like guided concretization

    What stands out is not a clean leap to a brand new theory. The more convincing reading is that the model found a concrete exploitable structure that prior work had already identified as sufficient for an attack if discovered. That puts this result in the same bucket as recent LLM math counterexamples, where the model is strong at searching for witnesses once the shape of the answer is known.

    Use LLM-based research workflows where success can be checked quickly and prior theory narrows the hunt. Expect more value from example-finding and search than from open-ended conceptual invention.

      Attribution:
    • Diogenesian #1
    • recitedropper #1
    • xmcp123 #1
  3. 03

    Human expertise and dead ends are underplayed

    The polished narrative hides how much domain knowledge and wasted exploration sat behind the final paper. Commenters pointed out that the people driving these projects already have strong math backgrounds, and the published successes tell you nothing about how many expensive runs went nowhere. That makes the result more like expert-augmented search than autonomous discovery in the simple sense.

    When evaluating AI R&D claims, ask for denominator data. Success stories without failed-run rates will make the capability look more mature and reliable than it is.

      Attribution:
    • wahern #1
    • influx #1
  4. 04

    Multi-agent gains may just be search diversity

    The most interesting part of the pairwise worker story is not “collaboration” in the human sense. It is that independent trajectories explore different branches, so one agent can keep working on an idea another discarded too early. That suggests the scaffold’s value may come from parallel stochastic exploration and replayability, not from deep synthetic teamwork.

    If you build agent systems, invest in parallel runs, branch recovery, and reproducibility metrics. Measure hit rates across reruns instead of treating one successful trace as evidence the workflow is robust.

      Attribution:
    • a-dub #1
    • imightbebatman #1
  5. 05

    Six-figure token budgets are a real moat

    Several people fixated on the reported $100,000 per result because it signals a capability gap, not just a research cost. Frontier labs can run long autonomous searches with high throughput, privileged model access, and custom scaffolds that outsiders cannot easily match. Some saw that as a temporary phase of a new compute-heavy technology, others as the start of an AI research aristocracy.

    Assume frontier-model research is ahead of what your team can reproduce from public APIs. Budgeting, rate limits, and harness quality now shape who gets to discover things first.

      Attribution:
    • mmaunder #1
    • jrflo #1
    • kmoser #1
  6. 06

    Prompting skill is mostly task steering

    The slapdash prompts in Anthropic’s own examples undercut the idea that magic wording is the core skill. The useful competence is knowing what outcome you want, spotting drift, and deciding when to intervene versus letting the model continue. Better prompts still buy efficiency, especially when token budgets are not infinite, but the durable skill looks a lot like managing a junior coworker with poor calibration.

    Train teams on evaluation and correction, not just prompt templates. The biggest gains come from choosing good tasks, catching subtle failure, and using tighter prompts only where the payoff is clear.

      Attribution:
    • _dwt #1
    • alwa #1
    • hgoel #1
    • thephyber #1

Against the grain

  1. 01

    Chosen-plaintext attacks are not obsolete

    The attempt to minimize the AES result by calling chosen-plaintext scenarios unrealistic goes too far. Encrypting oracles still exist in real systems, and implementation details can reopen attack surfaces that clean textbook arguments dismiss. Strong attack models remain relevant precisely because deployed systems are messy.

    Do not let “not practical today” harden into “never relevant.” Keep threat models broad when reviewing cryptographic designs and the software around them.

      Attribution:
    • vessenes #1
  2. 02

    Human researchers will not stop trying

    The fear that hard unsolved problems become psychologically off-limits once AI fails on them may be overstated. Mathematicians and cryptographers already work on problems because they are hard, and AI’s main benefit today is clearing away easier lines of attack so human attention can focus where judgment matters most.

    Treat AI failure as triage, not a verdict. In research settings, use it to eliminate cheap possibilities, then push human effort toward the stubborn remainder.

      Attribution:
    • some_furry #1
    • himata4113 #1

In plain english

AES ↩
Advanced Encryption Standard, the most widely used symmetric encryption algorithm for protecting data.
AES-128 ↩
A version of AES that uses a 128-bit key.
API ↩
Application Programming Interface, a way for software to send requests to another service or program.
automorphism ↩
A symmetry of a mathematical structure that maps the structure onto itself while preserving its form.
chosen-plaintext attack ↩
An attack model where the attacker can choose specific plaintexts to be encrypted and study the outputs.
cryptanalysis ↩
The study of finding weaknesses in cryptographic systems or proving how hard they are to break.
HAWK ↩
A proposed post-quantum digital signature scheme that has been considered in standardization work.
HAWK-256 ↩
A specific parameter set or variant of the HAWK signature scheme.
LLM ↩
Large language model, a type of AI system trained on huge amounts of text that can generate and analyze language and code.
scaffold ↩
The surrounding software harness that gives a model tools, memory, parallel workers, and task control.
signature scheme ↩
A cryptographic system used to prove that a message or software update really came from a specific signer and was not altered.
token ↩
A chunk of text a language model reads or generates, commonly used as the unit for pricing and limits.

Reference links

Cryptography references

Math and proof workflows

Tools and related technical references

  • cicada
    Code search tool mentioned as useful context for agent instructions in some repos
  • AsicBoost paper
    Example of algorithmic optimization for SHA-256 that commenters wanted AI to search for analogs of