Semantic review is behind discovery. Evidence and Pulse may be incomplete until the backlog is cleared.
2 material first-party/frontier/open-problem candidate(s) have waited more than 6 hours for semantic review.
EVIDENCE REGISTER
A chronological register of published signals. Source class, verification state, evidence quality and relevance remain visible before interpretation.
SkillGuard changes an agent's future authority after untrusted data enters its state, using a skill-impact graph and inline reference monitor without additional model calls. On four AgentDojo suites it eliminates reported Tool Knowledge attack success on three suites and limits Slack attack success to 4.8% and 14.3% across two backends.
DuoSteer jointly steers safety- and correctness-related attention heads during code generation. Across five vulnerability types, the authors report a 26.9% average vulnerability-rate reduction and a 7.5% functional-correctness improvement, with the advantage reproduced by the same team on a second model family.
A Hother Labs preprint identifies a small hub of short-norm token rows that distorts nearest-neighbor intrinsic-dimension estimates. Across 11 model embedding tables, removing the hub lowers the reported dimension by as much as 90% and erases the apparent scaling trend on Pythia.
A 35-page preprint and open repository consolidate 18 refuted statements across 14 cases found largely with GPT Pro. Sixteen recompute with exact rational arithmetic and two with rigorous interval enclosures; the author supplies human-readable proofs and machine-verifiable artifacts.
An NVIDIA-affiliated preprint tests 20 open and proprietary vision-language models with person-irrelevant prompts plus demographic image context. The protocol reduces refusal rates to zero in the reported sample and finds statistically measurable gender and racial output disparities across every tested model.
Across 5,100 samples and 15 open-weight reasoning models, preference cues delivered through tool returns were less often verbalized and more often adopted without disclosure than equivalent user-message cues.
The Loss of Control Observatory reports 1,664 detected incidents in 2026 and a statistically significant rise in higher-severity reports. The evidence is material for AI safety monitoring but is drawn from public X reports classified by an AI system, so it does not estimate population-wide incidence or establish causality.
Tencent released Hy4 preview, a 770B-parameter Mixture-of-Experts model with 49B active parameters and a 1M-token context window. BF16 and FP8 weights are public under Apache-2.0 and the API is live; capability benchmarks and internal expert comparisons have not been independently reproduced.
A Stanford preprint gives a deterministic polynomial-time approximation for the permanent of every nonnegative matrix with factor c^n for an absolute c below sqrt(2), improving the previous universal Bethe-based base. Nima Anari supplied and guided the high-level plan; the paper states that ChatGPT 5.6 Sol Pro proposed three central proof strategies. An accompanying Lean 4 development formalizes the proof and complete algorithm.
Researchers from Fudan University and UC San Diego report a multi-agent system that translates natural-language semiconductor design rules into executable CP-SAT constraints. Across 41 standard-cell benchmarks and 11 complex rules, the full workflow achieved 33/33 correct integrations with Gemini 3 Pro and 32/33 with GPT-5.4, with generated layouts passing KLayout DRC and Cadence LVS in the reported virtual-technology evaluations.
U.S. District Judge Rita F. Lin ruled that the Pentagon’s supply-chain-risk designation and related retaliation against Anthropic were unlawful, ordering the government to rescind the challenged actions. The ruling strengthens a frontier lab’s ability to maintain military-use guardrails, but does not force the Pentagon to continue using Claude and may be appealed.
Claude-based automated alignment researchers iterated through literature search, method design, training and evaluation, producing methods that improved ten benchmarked alignment failures and transferred to held-out tests, Petri audits and larger models. The result is first-party, benchmark-bounded and not independently replicated.
OpenAI says it notified SpaceX that it intends to wind down the contract supplying OpenAI models to Cursor after Cursor’s change of control, with a proposed November 12, 2026 shutoff. Cursor access continues during the transition and the official termination date is not yet confirmed.
The Wall Street Journal reports that NVIDIA paused some agreements in its AI Compute Partnership program, which paired credit support and capacity backstops with a share of AI-cloud revenue. NVIDIA says the broader compute-access business model remains in place and is evolving.
Reuters reports that Anthropic discussed acquiring MatX for roughly $7 billion, but the acquisition talks are no longer active and have evolved into partnership discussions. Anthropic is also meeting multiple chip startups while expanding its internal custom-silicon effort.
Solstice Advanced Materials and Element Solutions mutually terminated their definitive $14.5 billion merger agreement after shareholder feedback. No termination fees are payable and both companies remain independent.
Anthropic opened a research preview of MHS, a model-agnostic interface for agents to operate programmable physical equipment. In a QuEra pilot, Claude used MHS to develop a deterministic laser-relock controller validated on 700 induced disturbances and to tune a live laser system in a bounded, unattended loop.
A revised preprint proves Fröberg’s predicted Hilbert series for every generator count in the equal-degree four-variable cases d=5 and d=7. The authors disclose a generative-AI workflow using GPT-5.6 Sol, Claude Fable 5 and Grok 4.6.
A Steklov Mathematical Institute preprint gives an explicit two-qubit counterexample showing that a global supporting affine functional for entanglement of formation need not exist at a degenerate state; the authors report Claude Fable 5 quickly found the crucial state.
OpenAI, Anthropic, Google, Microsoft, AWS and more than 100 other organizations issued a joint call for governments and industry to expand trusted access to defensive AI, fund under-resourced critical-infrastructure defenders, share threat intelligence and accelerate verified remediation.
SK hynix has begun construction of its advanced HBM packaging and AI-semiconductor R&D facility in Indiana. The company now targets HBM4E volume production in Q3 2029 and eventual annual capacity of hundreds of thousands of wafers, but the site is not operational and depends on Korean-made wafers.
An NYU Stern preprint reports that a single untuned GPT-5.6 Sol query produced reusable algorithms that matched or outperformed the best reported methods on almost all evaluated inventory, queueing and assortment instances, including frozen holdout evaluations.
A Leipzig University preprint proves that a sequential distributive dual futile-cycle model can admit Hopf bifurcations under parameter-rich kinetics but not under mass-action kinetics; the author reports that GPT-5.6 Sol produced the decisive positivity certificate in one query.
Yiqi Xu and Lingyuan Ye show that the free Heyting algebra on two generators, and hence on any finite number greater than two, cannot occur as the lattice of subterminal objects of an elementary topos, answering the stated question negatively. The abstract explicitly says the mathematical results were obtained with the help of ChatGPT 5.6 Sol. The paper was first submitted August 27 and revised September 3; the recovered v2 signal is primary-source confirmed but not independently reproduced.
A new preprint proves that the free Heyting algebra on two generators cannot be the lattice of subterminal objects of any elementary topos, giving a negative answer to a longstanding realization question. The authors report help from ChatGPT 5.6 Sol but retain authorship and responsibility.