Superconductivity in graphene holds past the field that should end it
MIT physicist Long Ju and collaborators measured three superconducting states in pentalayer graphene that survive an in-plane field of up to about 8.5 tesla, tens of times the strength that usually breaks a superconductor.
A magnetic field enhances and induces superconducting states in pentalayer graphene1
A magnetic field normally destroys a superconductor by pulling apart its spin-paired electrons, and the field strength at which that happens is the Pauli limit. A Nature paper in the August 6 print issue reports three superconducting states in pentalayer rhombohedral graphene that outlast that limit. They survive an in-plane field of up to about 8.5 tesla, tens of times the Pauli bound for their transition temperature.12
Two of the three states do more than survive the field: it creates them. They appear only once the field is applied.2 Under a perpendicular field the transition temperature climbs from about 55 to about 90 millikelvin, still far too cold for any use outside a dilution refrigerator.2 The measurements sit in the clean limit, where disorder is low enough that the behavior reflects the material and not its defects. The result reached the web in early July; the collaboration runs through MIT, the University of Basel, Florida State, the University of Florida, and Japan's National Institute for Materials Science.2
OpenAI's ten Astra proofs are Lean-checkable and still unrefereed3
When this paper covered OpenAI's ten machine-checked proofs on August 4,
no one outside the company had read them. The first outside read has now
arrived, and it is informal. Thomas Bloom, the University of Manchester
mathematician who curates erdosproblems.com, called the results
“big news” while noting the system was built by
mathematicians and trained on human mathematics.4
That is a working mathematician's reaction, not a referee's: none of the
ten results has been through peer review.4
What can be checked today is the Lean. OpenAI published the proofs as
Lean 4 certificate files in a repository, openai/ten-proofs,
that anyone running Lean 4 can verify line by line.3
Lean shows each stated theorem follows from its axioms; it does not show
that referees have accepted the problems as solved.
The underlying claim, from OpenAI's August 1 announcement, is that an internal, unreleased version of a model it calls Astra generated solutions to ten problems in mathematics and theoretical computer science, each unsolved for at least a decade.3 OpenAI's own phrasing was careful — the results “resolve or make substantial progress on” the problems — and it credits Astra with the mathematical arguments while taking responsibility for formalizing them in Lean.3 One result, the existence of a non-sofic group, would close a question open since Gromov introduced soficity in 1999.4 The Decoder put the run's compute at about $2,000 at OpenAI's Sol API rates.4
GPT-4 forecasts social-science experiment results about as well as human panels5
A Nature paper posted online July 8 tested whether a language model can predict what a social-science experiment will find.5 Ashokkumar, Hewitt, Ghezae, and Willer prompted GPT-4 to simulate representative American samples for a large archive of survey experiments. Its predicted treatment effects correlated strongly with the measured ones, at accuracy comparable to pooled human forecasters.5 The correlation held even on studies published after the model's training cutoff.6
The model was not evenly calibrated: it systematically overstated how large the effects would be.6 A second archive of 15 megastudies and 606 effects gave lower correlations in the same range.6 That upward bias matters for planning: a study sized from an inflated forecast will be underpowered for the effect it actually finds.
CISA flags an actively exploited Langflow flaw that hands over an app's model keys7
On August 4 CISA added CVE-2026-9198, an unauthenticated remote-code-execution flaw in Langflow rated CVSS 9.8, to its Known Exploited Vulnerabilities catalog, with a federal remediation deadline of August 7.7 Langflow is a widely used builder for LLM agent workflows. The reach is the concern: a Langflow instance usually holds the model and connector credentials for everything it orchestrates.8
The exploit is a two-step chain. A request to
/api/v1/auto_login mints a superuser token for any caller
that can reach the server, and /api/v1/validate/code then
runs attacker-supplied Python through exec().8
IBM, which maintains Langflow, disclosed and patched the flaw in version
1.10.1 in July; versions 1.0.0 through 1.10.0 are affected. By the time
CISA catalogued it, exploitation was already live. One tracker logged
about 650 attempts against the flaw since July 6, from 244 addresses
across 41 countries.8
Sources
- Nature · Family of magnetic field-boosted superconductors in rhombohedral graphene
- Graphene-Info · Researchers find rhombohedral graphene can host multiple magnetic-field-boosted superconducting states
- OpenAI · Ten advances in mathematics and theoretical computer science
- The Decoder · OpenAI announces its next major model, Astra, by dropping ten previously unsolved math solutions
- Nature · Ashokkumar et al., Large language models can predict the results of social science experiments
- PubMed (NLM) · Large language models can predict the results of social science experiments (citation and abstract)
- CISA · CISA adds three known exploited vulnerabilities to catalog (August 4, 2026)
- The Hacker News · CISA flags Langflow RCE, Tomcat, and N-central flaws under active exploitation