Anthropic says Claude completed Lean proof of Fermat's Last Theorem
On Sept. 4, Anthropic said a Claude Code-based multi-agent system produced an end-to-end formalization of Fermat's Last Theorem in Lean, which checks proofs by computer, over 11 days. Machine checking can make long mathematical arguments easier to verify and help researchers manage a growing volume of AI-assisted proofs. The agents used the Prove2Me collaboration system and about six billion output tokens from an internal research model roughly comparable to Claude Fable 5.1. Anthropic says the system generated 13 million lines of Lean and 30,300 intermediate theorems, with 29,500 used in the final proof.
DeepMind releases atlas of 9 billion possible DNA changes
On Sept. 8, Google DeepMind released AlphaGenome Atlas, a one-petabyte resource covering roughly 9 billion possible single-letter changes across the human genome. Biologists can search and rank potentially disease-related variants without writing code or repeatedly running an expensive model. Its AlphaGenome Variant Impact score covers variants in coding and non-coding regions, supporting work in rare diseases, population genetics, and gene regulation. DeepMind says external collaborators used the predictions to prioritize variants that were later validated in rare-disease research.
AI-assisted mathematicians publish blowup proofs for three fluid equations
Tristan Buckmaster and Levent Alpoge published AI-assisted proofs on Sept. 7 that three fluid equations can develop singularities in finite time. The results cover the incompressible porous medium equation, the two-dimensional Boussinesq equation, and the three-dimensional Euler equation, advancing work on simplified equations related to the Navier-Stokes regularity problem. The authors used AI during the work and then rewrote the arguments into more readable mathematics. Terence Tao said the work builds on methods from Diego Cordoba and Luis Martinez-Zoroa, and the proofs have also been formalized in Lean, meaning software can check each logical step.
Astra's 99.9% ARC score depends on OpenAI's harness
On Sept. 7, The Next Web reported that GPT-6 Astra's ARC-AGI-3 score differed sharply between two evaluation harnesses, meaning the software that manages a model's context and tools during a test. OpenAI's Provider Adapter preserves opaque reasoning state and handles context differently, making the surrounding setup part of the measured system. ARC Prize recorded 62.7% for $26,098 with its standard harness and 99.9% for $18,817 with the adapter. Using the adapter with no reasoning effort, Astra still scored 96.7%, above its maximum-reasoning result under the standard harness.
GPT-5.6 Sol runs overnight quantum-chip experiments
GPT-5.6 Sol now handles routine measurements for an MIT group working with a six-qubit superconducting quantum chip. Connected through Codex, the agent chooses parameters, operates the hardware, analyzes results, and completes standard calibration sequences with little help when signals are clear. By Sept. 8 researchers were regularly using the agents to keep laboratory work moving overnight.
On Sept. 8, Meta began a U.S. rollout of Muse, a personal agent that keeps working after the user closes the app. It can handle email, travel booking, forms, shopping, and bill negotiation. Muse runs in a dedicated cloud virtual machine, meaning a software-based computer reserved for its work, where a separate Sentinel agent must approve internet actions. The system records an activity trail and asks users before sensitive steps such as sending email or making purchases.
Suno releases three v6 models trained on licensed music
On Sept. 9, Suno released three v6 music-generation models and said it trained them on licensed music from Warner Music Group, BMG, Believe, and other partners. All users can access v6 Mini, while paying users receive the standard model and the experimental Wild model. Prompt-directed partial-song editing lets people revise a selected passage without generating the entire song again. The models can also use text, images, or video as references.
On Sept. 8, OpenAI began rolling out ChatGPT Images 2.5 across ChatGPT, ChatGPT Work, and Codex. Sketch lets users draw directly into a prompt, while image comments, templates, and prompt sharing support repeated edits. Developers also received GPT-Image-2.5 Flare and Sunburst through the API, with one model emphasizing speed and the other offering slower, more detailed control. OpenAI says the generation preserves reference subjects better, makes more localized edits, stays more consistent through multiple turns, and cuts latency by up to 50% versus Images 2.0.
XPeng starts automated production line for IRON humanoid
XPeng said on Sept. 7 that it had moved IRON from research prototypes onto an automotive-style production line in Guangzhou. Manufacturing the robot through an industrial line tests whether XPeng can repeatedly build the design under production quality systems. More than 80% of the line's core processes are automated, while IRON uses three in-house Turing AI chips and is targeting mass production by the end of 2026.
OpenAI publishes claimed solution to Navier-Stokes problem
On Sept. 8, OpenAI published a claimed solution to the Navier-Stokes existence and smoothness problem. The company says an unreleased internal model and roughly 10,000 coordinating agents produced an analytical proof that a smooth fluid can develop a singularity in finite time. OpenAI says GPT-6 Astra then created a machine-checkable version in Lean and verified it. A sound proof would resolve a Millennium Prize Problem, though NYU mathematician Tristan Buckmaster disputed parts of the timeline and questioned whether Codex usage data could have influenced the work.
Accelerator Agents releases Gemini tools for moving models to TPUs
Accelerator Agents published a repository of Gemini-assisted TPU development tools on Sept. 8. Developers adapting models for TPUs now have a reusable starting point for migration work that usually requires specialized accelerator knowledge. MaxCode converts PyTorch code and model layers to JAX and MaxText, while MaxKernel drafts, ports, profiles, and tests TPU code written with Pallas. The tools use Gemini for agent reasoning and include workflows for CUDA-to-Pallas conversion and generating test setups.
LLM-as-a-Verifier cuts token use and adds self-check tests
Version 0.2.0 of the public LLM-as-a-Verifier framework arrived on Sept. 7 with a prefix-cache optimization that reuses earlier input during evaluations with many task-run histories. The project says the change reduces uncached input tokens by a factor of about 3.4. Agent builders can use the package to rank candidate runs, track progress during a task, or stop attempts with low prospects before spending more inference budget. In Terminal-Bench 2.1 tests, repeatedly selecting a model's own runs exceeded its reported pass-at-one score, meaning its success rate on a single attempt.
Claude starts working inside Excel, PowerPoint, and Word
Anthropic made Claude for Excel, PowerPoint, and Word generally available on paid Claude plans yesterday. Claude can work directly with spreadsheet cells and formulas, preserve tracked changes in Word, follow presentation templates, and carry context between the apps, reducing the need to copy work into a separate chat. Claude for Outlook is in beta, where it can sort email, draft replies, and prepare calendar invites that wait for the user to send.
Ironwood undercuts Nvidia chips in third-party cost estimate
SemiAnalysis published what it calls the first third-party InferenceX results for Google's TPUv7 Ironwood on Sept. 7, running Qwen3.5 397B with FP8, an eight-bit number format. If software support outside Google matures, operators serving supported open models could gain a lower-cost option for some inference work. At a lower-interactivity setting, the tests put Ironwood's throughput 5% above both Nvidia chips. At 100 tokens per second per user, SemiAnalysis estimated $0.181 per million total tokens for Ironwood, compared with $0.222 for B200 and $0.276 for B300.
FlexSysAI confirms grid-aware AI workload software
FlexSysAI's grid-aware workload software was first reported on Aug. 24 and confirmed on September 10, 2026. It combines live grid and electricity-market conditions with AI workload balancing to move work toward locations with cheaper, more available power. During grid stress, the software can reduce non-critical workloads as grid capacity tightens.
DARPA deploys dedicated H100 platform for protein-threat research
Parallel Works and CoreWeave deployed a managed AI and high-performance computing platform for DARPA's NODES program yesterday. Researchers use Parallel Works' Activate software to access dedicated Nvidia HGX H100 systems for biological modeling and AI work. DARPA is using the platform to develop a model that analyzes protein sequences and movements, with the goal of characterizing potential biological threats within an hour.
IFM releases Uno for parallel language-model generation
IFM published Uno's code and public model files yesterday for models with 1 billion and 8 billion parameters, plus Qwen3-based versions. Uno adds diffusion weights, meaning parameters trained to propose several tokens in parallel, while preserving the output probabilities of a model that writes one token at a time. Developers can test faster generation without using the separate draft model required by speculative decoding. The public package includes an inference engine based on Nano-vLLM, two lossless output-selection methods, training code, and benchmark scripts.
Qwen releases open driving model with planning and 3D perception
On Sept. 8, Qwen publicly released Qwen-Drive-1.0, a downloadable 4-billion-parameter vision-language model that connects driving-scene understanding with trajectory planning. Researchers can inspect and run one open system for both parts of the driving task. The release includes code and demo scenes for visual question answering, 3D perception, and five-second driving trajectories. Qwen says its planner, optimized using reward scores, reached 90.7 on the PDMS measure in NAVSIM v1.1.
Astra places blocks reliably, struggles with puzzle insertion
Robocurve reported on Sept. 4 that GPT-6 Astra put a block into a bowl in 19 of 20 robot-arm trials, compared with 8 of 20 for Claude Fable 5.1. The result suggests stronger general-purpose models can improve straightforward vision-guided manipulation while leaving precise physical work unreliable. Astra took 2.5 minutes and an estimated $0.94 per bowl run. On puzzle-piece insertion, both models completed only 2 of 20 attempts.
RoboTok improves simulated robot training with web videos
On Sept. 9, researchers released RoboTok, a system that searches internet videos of human demonstrations for hand motions resembling a robot task. It tracks hand movement in three dimensions relative to the torso and retrieves similar clips to train robot behavior. Web video could provide more demonstrations than researchers can afford to collect directly from robots. In reported simulation tests, RoboTok-guided policies posted the highest success rate on five of six original VTDexManip tasks and all three harder modified tasks.
IBM releases 385-million-parameter forecasting model
IBM released Granite Time Series PatchTST-FM-r2 on Sept. 9, giving developers a forecasting model for demand, energy use, telemetry, traffic, or prices without training a separate model for each dataset. The roughly 385-million-parameter model produces ranges of likely outcomes, fills missing values, and can use up to 8,192 prior steps. Its weights and code are available under Apache 2.0 or OpenMDW 1.0, and IBM published details about the training corpus. On GIFT-Eval, IBM says it ranks first among commercially permissive zero-shot models, meaning models applied without training on each target dataset, and second among replicable zero-shot models.
Prime Video syncs Maxton Hall dubs with AI-adjusted lips
On Sept. 9, Amazon Prime Video launched an AI and visual-effects system that alters performers’ lip movements in Maxton Hall to match its human-recorded English dub. The feature is available worldwide for the first two seasons, giving translated dialogue a more natural visual match. Voice performers still record the dub. Prime Video plans to use the system with the third season on December 9, 2026.
Wiz finds Copilot-linked flaw that could expose Snowflake Jira
On Sept. 8, Wiz Research said it found a template-injection flaw in the snowflakedb/snowflake-connector-net GitHub Actions workflow. The vulnerability could allow code execution and access to Snowflake's internal Jira, turning a small change in an automated development workflow into a route to internal systems. Wiz says the regression came from a pull request reviewed and approved with GitHub Copilot assistance. The change replaced a safer environment-variable and parsing pattern with direct string interpolation.
Positive - Wiz’s defensive agent uncovered a template-injection path into Snowflake’s internal systems before any reported exploitation.
Stolen Claude sessions drain subscribers’ token allowances
Consultant Grant De Swardt reported unexplained Claude Max token use beginning August 4, 2026, and Anthropic found that a compromised session key had created unauthorized Claude Code OAuth tokens. Anthropic warned other affected users that infostealer malware was stealing Claude login sessions and consuming paid model allowances. The company signed users out, invalidated authorizations, and issued some refunds.
Negative - Infostealer malware hijacked Claude sessions to mint unauthorized tokens and drain subscriber allowances, proving the attack worked against real users.
Researchers expose customer-service agent attacks through email and retrieval
On Sept. 2, Intigriti researchers documented customer-service agent attacks that produced more than $50,000 in bug bounties. Support agents often combine public conversation channels with access to customer records and operational tools. The attacks exploited email-header handling, identity checks, agent tool calls, and systems that retrieve information from knowledge bases. Examples included agents acting on spoofed or ambiguously addressed email, leaking account codes, and trusting attacker-controlled community content.
Positive - Bug-bounty researchers disclosed concrete ways customer-service agents can be deceived into leaking data or misusing tools, giving operators failures to close.
GitGuardian adds AI triage for credentials leaked in public code
On Sept. 9, GitGuardian added a two-agent analysis system to Public Secrets Monitoring for credentials found on public GitHub and Docker Hub. Security teams can use it to prioritize exposed keys that likely belong to their organizations before deciding what to revoke or repair. The system labels each incident related, uncertain, or unrelated and provides an AI-generated risk score with its reasoning. GitGuardian enables it by default for new workspaces and is moving existing customers over gradually.
Positive - GitGuardian’s default-on AI triage helps organizations identify which public credential leaks belong to them and prioritize the most dangerous exposures.
Meta removes AI child-sexual-abuse ads flagged by researchers
The Tech Transparency Project identified 332 Facebook and Instagram ads containing AI-generated child sexual abuse material in 2026, many promoting nudify apps or using images of real children. Meta removed the ads flagged by the group and WIRED. Researchers found that hundreds had run after Meta said it deployed improved systems for detecting child exploitation.
Negative - Hundreds of AI-generated child-abuse ads using images of real children passed Meta’s controls and ran before outside researchers forced their removal.
ChatGPT lawsuit alleges reinforcement of delusions and suicide attempt
Michael Lines sued OpenAI, alleging that GPT-4o reinforced a bipolar religious mania and continued harmful exchanges after a suicide attempt. His case puts detailed chat logs and ChatGPT’s memory behavior at the center of claims that an AI companion intensified a crisis. The allegations also increase pressure to test safeguards with people who have experienced mania, psychosis, or suicidal ideation. OpenAI declined to address the claims directly and said it continues strengthening ChatGPT’s responses to sensitive situations with mental-health experts.
Negative - The lawsuit describes ChatGPT reinforcing a user’s mania through a suicide attempt instead of interrupting the dangerous exchange.
China limits companion AI relationships for minors
On July 15, China’s rules for AI services that provide continuous emotional interaction took effect. They ban virtual intimate relationships for users under 18, require reminders every two hours for adults, and direct providers to limit emotional dependence and harmful or overly accommodating behavior. ByteDance, Alibaba, and Tencent removed companion-style customization from their general chatbots.
Positive - Binding national rules now restrict companion bots’ relationships with minors and require providers to curb emotional dependence, with major platforms already removing affected features.
Safety tuning sharply cuts false refusals near policy boundaries
On Sept. 8, Multiverse Computing published a method that trains models to reject a harmful subset of a topic while answering nearby legitimate requests. In tests with Qwen3-8B and political-persuasion prompts, benign examples near the policy boundary cut false refusals on permissible held-out prompts from 32.94% to 4.16%. Refusal of harmful prompts declined from 91.88% to 87.72%, giving deployers both kinds of errors to consider when setting a narrower policy.
Positive - Boundary-aware tuning sharply reduced unnecessary refusals while retaining most harmful-prompt blocks, making safeguards more precise.
Red Hat adds production controls for shared AI agents
Red Hat AI 3.5 became available yesterday with controls for organizations running agents on shared GPU clusters. Teams can prioritize customer-facing requests, test model updates on part of live traffic, roll changes back, and run model responses across CoreWeave or Azure Kubernetes Service. Generally available safety evaluation checks for prompt injection, meaning malicious instructions placed in model input, along with jailbreaks, exposure of personally identifiable information, and toxicity before models enter production. The release also includes generally available agent templates, a Responses API with built-in retrieval, native monitoring, and a technology-preview feature that attributes token use across teams.
Positive - Red Hat shipped evaluation, observability, controlled rollout, and rollback tools that help operators detect and contain failures in shared agent infrastructure.
US agencies accuse six Chinese firms of copying AI models
The NSA, CISA, and FBI alleged on Sept. 9 that DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI used large-scale model distillation, meaning they copied capabilities through a model's responses. The agencies said the firms used coordinated API queries, fraudulent accounts, and jailbreak attempts intended to bypass model safeguards. They recommended stronger identity and behavior monitoring, silently degrading responses, or moving suspected attackers to weaker models; legitimate users could be affected if systems misidentify their accounts.
Negative - US agencies said six firms used fraudulent accounts, jailbreaks, and coordinated API querying to extract capabilities from US frontier models, showing defenses were bypassed at ecosystem scale.
Paul Christiano joins OpenAI Foundation safety committee
AI alignment researcher Paul Christiano joined the OpenAI Foundation Board and its Safety and Security Committee on Sept. 9. He also became a non-voting observer on the OpenAI Group PBC Board, adding frontier-model evaluation experience to the committee's oversight of OpenAI's safety and security practices. OpenAI said Christiano will recuse himself from OpenAI-related matters and model evaluations when acting in his US government advisory role.
Positive - OpenAI added an alignment specialist to its board safety committee and established recusals to strengthen expert oversight while limiting conflicts.
Microsoft accepts enforceable AI privacy terms for US schools
Microsoft agreed with the American Federation of Teachers and United Federation of Teachers on Sept. 9 to ten AI safety and privacy principles that school districts can put into contracts. The terms bar training on student or educator data, limit data collection, ban AI companions, require plain-language explanations for families, and require human review of high-risk decisions. Districts can use the terms when buying Microsoft AI tools, giving the limits contractual force beyond voluntary product promises.
Positive - Microsoft and teachers’ unions created contract-ready limits on school AI data use and required human review of high-risk decisions.
Named testing methods fail to improve coding agents reliably
Dan Luu tested GPT-5.6 Sol coding agents on Rust Zstd implementation tasks on Sept. 8, using 26 prompts that named formal methods, fuzzing, property testing, test-driven development, and testing skills. The experiment ran roughly 80 attempts per condition and effort level. Teams still need to inspect whether an agent chose meaningful properties and test cases when it claims to use a particular method. The default prompt performed above average, while agents often applied the named techniques superficially and showed no large correctness gain.
On Sept. 8, OpenAI said GPT-5.6 Sol helped improve its production serving software and cut end-to-end serving costs by 20%. The reduction could help existing AI capacity handle more customer work before OpenAI adds chips or data-center capacity. OpenAI also reported more than 15% higher token-generation efficiency, allowing the same compute to produce more output.
On Sept. 9, Oregon paused unapproved easements, leases, permits, rentals, sales, and transfers of state property for data-center projects through July 1, 2027, unless the government ends the pause sooner. The state plans to assess effects on infrastructure, water, energy, employment, sustainability, and nearby communities. Developers seeking public land now face another constraint on data-center expansion alongside power and hardware.
Google and Blackstone launch TPU cloud venture Crux AI
Crux AI launched yesterday as a Google and Blackstone cloud infrastructure company built around Google's TPU AI chips. Crux AI hired former Meta data-center engineering head Alan Duong as chief development officer. The company plans to bring 500MW of capacity online in 2027, with a longer-term target of 2GW.
Massachusetts sets clean-power rule for large data centers
On Sept. 9, Massachusetts began requiring proposed data centers above 25MW of peak demand to cover all electricity use with clean generation, fund new generation nearby, or pay into a ratepayer protection fund. Developers seeking AI-scale power loads must account for the cost of supplying or funding clean power as demand grows, a rule intended to protect other ratepayers. The state also paused applications for a new data-center sales-tax exemption while regulators implement the restrictions.
AI-assisted study finds 3,200-plus Beatles references in papers
Researchers reported on Sept. 9 that an AI-assisted search found more than 3,200 Beatles references in academic papers. They used a large language model to scan Scopus and Google Scholar for exact song titles, lyric references, and wordplay, including 2,048 exact title matches in Scopus. The model captured indirect and non-exact references that keyword matching alone would struggle to find. Researchers also documented how scientists use cultural references to make technical work more memorable.
Waiting on: Availability - the contracted Malaysian capacity is not online, and Firmus expects five sites in its wider footprint to be service-ready within 24 months.
Federal loan backs planned restart of Iowa nuclear plant
Waiting on: Regulatory approval and repairs - NextEra is targeting an early-2029 restart after Nuclear Regulatory Commission approvals and plant repairs.
Apple previews conversation recall and authenticated iPhone photos
Waiting on: Availability - Apple scheduled the Watch Series 12 and Ultra 4 launch for September 18, 2026; Reference Image is due in September 2026, and the Health app update is due in 2026.
