ma2tic

The brief · 29 July 2026

The AI brief, 2026-07-29

Science is letting the agents off the leash: Claude is hunting down crypto vulnerabilities, entire labs are coding with AI, and materials are being deciphered without ever being broken. Which leaves the one question that stings: when a model forgets something, does it actually stay forgotten?

3 min read J / K to navigate

Claude finds mathematical flaws in encryption algorithms on its own

Until now AI could spot coding bugs in cryptography libraries, not flaws in the algorithms themselves. This time Claude improved the best known attack against a post-quantum signature scheme that had already passed two years of expert review. Nothing in production is broken yet, but the playing field for cryptography research has just shifted.

practitioners › HAWK, a candidate in NIST's post-quantum competition, was attacked by Claude Mythos Preview in 60 hours, cutting its key strength in half.

www.anthropic.com →

Genomics researchers hand their code over to AI agents

Science teams are using coding agents to modernize computational tools that are sometimes twenty years old, without rewriting all the legacy code by hand. For a lab with no dedicated dev budget, this changes how fast an idea becomes a usable pipeline.

practitioners › Documented use case in genomics, reported gains on code porting time, no published performance figures.

Field report scientifique →

Erasing information from an AI model does not stay erased

A company that certifies today that a piece of data has been removed from a model cannot guarantee that removal still holds after the next update. For any use case bound by GDPR or a right to erasure, compliance achieved on an AI model is a snapshot, not a permanent state.

practitioners › Tested on the reference implementation AI Engram (Kwon et al., 2026), three models from two providers, gap of 61 to 71 percent between theoretical and actual edits after successive cuts.

arxiv.org →

More electrodes does not mean a faster brain

A theoretical paper challenges the assumption that adding more sensors to a brain-machine interface speeds up thought in equal measure. The body, the learning process and the need to confirm an action before executing it all cap the real gain, well before hardware limits come into play.

practitioners › Perspective paper, arXiv q-bio.NC, submitted July 17, 2026, no experiments run, capacity to gain relationship judged nonlinear.

arXiv →

An audit exposes false positives in simulated brain models

Researchers show a model can reproduce an observed brain signal without its parameters reflecting any real physiological mechanism. For epilepsy research and EEG imaging, this means some previously published conclusions may need a second look before being taken at face value.

practitioners › Tested on iEEG epilepsy data and EEG ERP CORE, the framework is called NMM-SBI Audit, published July 27 on arXiv (q-bio.QM).

arXiv →