ResearchNews
Claude wrote formal proofs for 123 open-source libraries
Anthropic and Theorem had Claude prove 239 entry points crash-free at the RISC-V binary level. What was proved, what it assumes and where it falls short.
Canberk / Archive
Research
Papers, experiments and new methods.News and analysis on this topic, together in one place.
18 articles
Home ↗
ResearchNews
Anthropic says Claude beat published quantum circuit designs
Anthropic says Claude's closed-loop designs beat published quantum circuits: 17x fewer gates for point addition, 12x for a DQI decoder. Claims, limits.
ResearchNews
How Claude traced NIST's ECDSA curve seeds to timestamps
Anthropic says Claude traced NIST's ECDSA seeds to lrand48 and Unix timestamps in under 30 minutes. What it solved, what it did not, how we checked.
ResearchNews
Can AI models self-replicate without a datacenter?
Anthropic says a 9B model set up inference on 8 of 9 hardware types. What its red team tested, the botnet estimate and the limits of the study.
ResearchNews
Anthropic distillation defenses: what the 90 to 99% claim means
Anthropic says new safeguards cut data harvested by distillers 90 to 99%. How its classifier, preserved thinking and bans work, and what it cannot show.
ResearchNews
Can AI review code without humans? CRBugBench finds the best agent catches 71.9% of known bugs
A benchmark of 1,800 code review tasks: the best agent catches 71.9% of known bugs in its top five findings, and less on proprietary code.
ResearchNews
METR's time-horizon plot, re-fitted: a Berkeley paper finds a flat stretch from 2 to 30 minutes
Two Berkeley statisticians refit METR's 228-task data. Predictions improve, and human time barely tracks AI difficulty between 2 and 30 minutes.
ResearchNews
Can you trust coding benchmark scores? TestJack's audit refuted a third of passing trials
A preprint re-audited 4,487 coding trials that passed benchmark tests and refuted 34.4%, cutting the resolution rate from 50.6% to 33.2%. It has caveats.
ResearchNews
Do sub-agents help coding agents? A 2,124-run study of Codex, Claude Code and Kimi Code
A preprint ran 2,124 coding tasks with sub-agents on and off: Claude Code lost 24 points on SWE-bench Verified and token use rose 1.4 to 3.3 times.
ResearchNews
Is auto-approve safe? What three new papers say about coding-agent security
A preprint found auto-approve lifted coding-agent attack success from 29.2% to 95.6% with one model. Two more show model and tool names move scores.
ResearchNews
EPFL and Apple study: a minimal coding agent matches elaborate ML harnesses
EPFL and Apple researchers report one plain coding-agent session matched four elaborate ML harnesses, though it cost about six times more than one of them.
ResearchNews
Red Hat: Jev is competitive with guardrail baselines, not clearly better
Red Hat tested Jev, Laya and DiffusionGemma against classifiers and LLM judges on two guardrail tasks. Jev ranked first on one and fourth on the other.
ResearchNews
Mythos found a login bypass in Rejetto HFS; attacks followed in a day
Horizon3 says Anthropic's Mythos chained a weak random-number generator into an admin login bypass in Rejetto HFS (CVE-2026-61500). The facts and limits.
ResearchNews
Anthropic's GLM-5.3 cyber report: what the tests show, and what they can't
Anthropic says GLM-5.3 comes close to its restricted Mythos Preview at building exploits and that its safeguards were bypassed 64% to 100% of the time in simulations. The figures, NIST's view, Z.ai's claims and the limits.
ResearchNews
Sakana AI SAIL: a robot tries the movement in simulation before acting
SAIL refines robot trajectories through simulation. Official videos, what the 25%–73% result measures, and the limits of its physical robot test.
ResearchNews
Google's four-part research path toward coherent long videos
Co-Director, CANVAS, A²RD and VQQA tackle different failure points in AI video storytelling. What the ten-minute demo and research benchmarks actually show.
ResearchNews
Claude and the ART enzyme finding: a research lead, with important unanswered questions
Anthropic's ART announcement explained: what Claude agents identified, how researchers followed it up, and why the biological function remains an open question.
ResearchNews
Google shares a library of AI-generated STEM learning activities
Google's new learning interactives combine AI generation with teacher review. The sample library and custom-generation pilot have different access paths.
Choose All to see other articles on this topic.