Cryptanalysis, scientific agents, and the gaps AI still leaves (2026)

The short read

The strongest posts in this window point to the same constraint: AI systems are getting better at execution, while validation, ownership, and unfinished human work remain stubbornly important. Four original posts made the cut.

Research: Anthropic says Claude Mythos Preview found improved attacks on HAWK and a reduced-round version of AES, while stressing that neither result breaks a deployed production cipher. 1
Scientific computing: OpenAI documents eight case studies in which coding agents handled work from maintenance to redesign, but scientists still had to define tests and judge the results. 2
Publishing: Ethan Mollick relays a study claiming that books without AI involvement earned less per book than in 2023 in seven of eight genres. The post does not identify the study, so the figure remains an attributed claim rather than an independently checked result. 3
Capability gaps: Mollick's separate point about "jaggedness" is more operational: an AI-assisted result can be substantially complete while still needing a human to close important gaps. 4

Cryptanalysis gets a serious test case

Anthropic's July 28 thread says Claude Mythos Preview helped researchers find weaknesses in two cryptographic targets. The first was HAWK, a post-quantum digital signature scheme that has not been deployed. Anthropic says the attack cut HAWK's effective key strength in half; for HAWK-256, the estimated work fell from 2^64 to 2^38. That is still an exponential attack, not a practical break of a deployed system. 5 1

The second result targeted seven-round reduced AES-128. Anthropic says Mythos Preview improved the best known attack by 200 to 800 times, but the result does not break full ten-round AES-128. Both projects cost roughly $100,000 in API usage, according to Anthropic's post. 6 7

What happened: A frontier model produced novel cryptanalysis against a candidate signature scheme and a reduced version of a widely studied cipher.
Why it matters: The useful change is in the research workflow: models can search for weaknesses that would otherwise require highly specialized human time.
What to keep straight: HAWK is not deployed, and the AES finding is not an attack on full AES-128. The immediate lesson is about standards review and defensive testing, not a new break of internet encryption.

Loading content card…

Anthropic also announced CryptanalysisBench with researchers at ETH Zurich, Tel Aviv University, and the University of Haifa. The accompanying paper describes 191 tasks across six families of cryptographic primitives, with separate tiers for known breaks, full-strength or scaled-down schemes, and frontier challenges. 8

Coding agents move the work upstream

OpenAI's field report covers eight scientific-computing case studies, including maintenance, targeted optimization, large-scale language migrations, GPU-native redesigns, and genomic software. In one example, an abandoned alignment project moved under new community stewardship; in another, a genomic parser was modernized. 2

The report's boundary is more useful than its sales pitch. Agents can execute ambitious, well-scoped engineering work, but they cannot reliably decide whether the result is scientifically valid. Researchers still need to define the question, specify how correctness will be measured, verify outputs against references or test targets, and decide who owns maintenance. 9

What happened: Eight case studies show coding agents taking on work that used to be blocked by implementation cost.
Why it matters: The scarce skill shifts toward choosing the right problem and setting a test that an agent cannot talk its way around.
The concrete risk: Cheap rewrites can fragment a project's maintainers. A generated replacement still needs a named owner and a credible maintenance plan.

Loading content card…

The book-market signal comes with a missing link

Ethan Mollick posted that a study found a flood of AI books is crowding out human authors. His quoted figure says books made without AI earned less per book than in 2023 in seven of eight genres; fantasy and horror were the exception, with a reported 35 percent increase. 3

This is a market signal, not a settled estimate. Mollick's post does not name the study, link to it, or provide its sample, definitions, or time window. The defensible takeaway is narrower: if the claim holds up, the pressure on human authors is showing up in earnings before it shows up in a clean replacement story.

What happened: Mollick relayed a study with a seven-of-eight genre result and a 35 percent fantasy/horror exception.
Why it matters: The relevant comparison is income per book, not the number of books published. More AI output can change the market even when readers still prefer some human work.
Confidence limit: The underlying study is not available from the post itself, so the figures should be checked before being used as evidence of a broad publishing trend.

Loading content card…

"Jaggedness" is the handoff problem

In a separate post, Mollick argues that the first AI-generated New York Times bestseller may have already happened in substance, even if it did not count technically because a human still had to fill the model's gaps. He adds math proofs as another example of work that can look nearly finished while retaining a decisive human step. 4

That is a more practical frame than asking whether a task is "AI-generated." The question for a team is where the remaining gap sits: fact checking, proof, taste, legal responsibility, or the final integration into a real workflow.

What happened: A result can be substantially produced by AI without being technically complete.
Why it matters: The last human contribution may be small in volume but decisive in whether the work is publishable, provable, or safe to ship.
Practical implication: Track the handoff points. They are where quality control and accountability still live.

Loading content card…

Cryptanalysis, scientific agents, and the gaps AI still leaves

The short read

Cryptanalysis gets a serious test case

Coding agents move the work upstream

The book-market signal comes with a missing link

"Jaggedness" is the handoff problem

References

Related content