AI News for September 15, 2026

Mozilla's State of Open Source AI report finds the capability gap between US closed frontier models and top open-weights Chinese models has narrowed to about 4.4 months, meaning companies pay roughly 5x the per-task cost for a four-month head start on tasks in the 8-to-12-hour range.

Salesforce unveiled Koa, its first reasoning model, post-trained with Nvidia on the open-weight Nemotron base to handle sales, marketing, and customer-support tasks as a cheaper, more secure alternative to frontier labs' models within its Agentforce platform.

Former Google DeepMind researcher Bilal Chughtai publicly warned that AI could 'kill us all' after resigning in July 2026, adding to a growing wave of safety warnings from industry figures like Anthropic's Evan Hubinger and Dario Amodei.

President Trump dismissed AI existential risk warnings as a 'hoax' while major AI CEOs urged a development slowdown, sparking debate over regulatory capture and genuine safety concerns.

Salesforce unveiled AIforce at Dreamforce, an AI interface layer with Claudeforce, Slackforce and Agentforce Coworker that embeds Salesforce business logic, security and governance into third-party AI agent workflows.

Anthropic's Claude Help Center release notes catalog product and model launches from July through September 2026, including Claude Fable 5.1 and Mythos 5.1, Opus 5, Salesforce integration, smart reports, memory updates, and Cowork expansion to web and mobile.

Salesforce's Dreamforce 2026 keynote introduced its first reasoning model Koa, a sweeping Customer 360 update, and an expanded Anthropic partnership, positioning the company around agentic enterprise AI.

An arXiv study of over 10,000 machine learning papers finds that repurposing existing datasets is associated with more disruptive research outcomes, especially when the repurposed data is later adopted by other researchers, and that such teams tend to be larger, more prestigious, and span academia and industry.

The PrimeScientist paper frames strategic allocation of research effort as a sequential decision problem, using an adaptive MCTS-based policy that improves average reward by 10.3% with 50.6% fewer research attempts than AutoResearch across 12 AI research tasks.

A tutorial-style guide surveying six open-source AI model families—Llama 3, Gemma, Mistral AI, Stable Diffusion, Qwen, and DeepSeek—with practical code examples for developers choosing foundations for coding and generation tasks in 2026.

A reference page tracking every Anthropic Claude model release from March 2024 through September 2026, with specs, pricing, benchmarks, and practical guidance for choosing among active models like Sonnet 5, Opus 5, Fable 5.1, and Haiku 4.5.

Elon Musk outlined a roadmap for Grok 4.7 through Grok 5, saying unreleased Grok 4.8 is a 2.5-trillion-parameter model nearing the end of training, though the claims remain unverified forward-looking statements.

President Trump publicly rejected frontier-lab calls to slow AI development, calling safety concerns a sick conspiracy and framing the race as one the US must win against China.

Congress faces a deadlock on AI regulation as Democratic lawmakers push for action before recess while President Trump and Republican leaders resist new guardrails and favor self-regulation.

The New York Times' copyright lawsuit against OpenAI and Microsoft is entering a pivotal phase as Judge Sidney Stein weighs whether the fair-use dispute will go to trial in 2027, with major implications for AI training and publisher economics.

The Manhattan District Attorney's Office seized 12 websites selling non-consensual AI-generated deepfake videos of roughly 1,200 people, marking a high-profile use of state law against harmful generative AI output.

Following warnings from Anthropic and other AI leaders about existential risk, lawmakers are pushing several competing AI regulation bills—including the Frontier Act, AI Kill Switch Act, and a proposed superintelligence ban—though near-term action looks unlikely before the midterm recess.

Also reported by: wtae.com

A Brookings research paper argues that AI exposure is the wrong basis for workforce policy, proposing targeted measures like wage insurance, apprenticeships, and domain-specific training to help displaced workers while expanding access to new high-return opportunities.

A U.S. magistrate judge will remain on the consolidated OpenAI and Microsoft copyright litigation after selling her Microsoft stock and concluding it did not influence her decisions.

A regularly updated U.S. map tracking copyright lawsuits against AI companies now lists 142 cases, including new suits against Microsoft and OpenAI.

Semafor opinion piece argues that AI companies are playing sophisticated defense in a broken Washington by shaping regulation rather than opposing it, including endorsing state-level rules in what OpenAI's Chris Lehane calls 'reverse federalism.'

AlleyWatch's daily funding report covers NYC startup rounds led by Profound's $180M Series D for AI search visibility and Fin.com's $20M seed for cross-border money movement.

A weekly VC funding roundup reports 74 rounds and $23.19 billion raised, down 37% week over week, with AI infrastructure and deep tech mega-rounds dominating while early-stage applied-AI deals stayed steady.

A daily startup funding roundup highlights more than €200 million raised by Dutch AI chip startup EUCLYD alongside a cluster of large Chinese rounds in robotics, semiconductors, sovereign AI infrastructure, and AI-driven drug discovery, signaling a shift toward AI's physical and infrastructure layers.

A venture capital roundup details over $745 million in new funding across AI infrastructure, coding agents, and networking startups, highlighting investor focus on deployment bottlenecks and defensible infrastructure.

LLRX's semi-monthly column compiles news, government documents, and reports on AI's impact on banking and finance, including Anthropic's Claude Money feature, AI boom financing risks, and AI skills in hiring.

A Wall Street Journal venture-capital analysis shows AI startup first financings surged from about 2,150 in 2022 to nearly 3,400 in 2023, but the middle of the market is now squeezed as investors grow selective and demand distribution, proprietary data, and defensible margins.

A retail investor explains his defensive strategy of buying healthcare stocks like JNJ, Merck, Eli Lilly, or the XLV ETF as a way to gain AI upside without direct exposure to AI-hype-dependent companies.

Fintech and bank executives at the Canada Fintech Forum say AI adoption in finance is real but scaling to millions of customers faces cost, security, traceability, and accountability hurdles.

IDC analysis shows agentic AI entering enterprise production in 2026, with token consumption projected to grow over 3,500% CAGR as agentic workflows—not agent counts—become the key AI infrastructure cost and planning metric.

A round-up of contact center AI moves from NiCE, Observe.AI, and Dialpad shows vendors shifting from AI claims toward proving workflow depth, governance, and measurable outcomes.

A reference changelog detailing Ollama's releases from version 0.32.0 through 0.34.2, including model support, engine improvements, and desktop app integrations.

FutureVault launched AI Agents that execute multi-step document workflows across banking, financial services, and insurance firms with built-in governance, permissioning, and audit trails.

A Google Public Sector and Accenture sponsored article describes how defense organizations like the Defense Logistics Agency are adopting agentic AI for logistics, agile operations, and administrative workflows.

Statistics Canada data shows wholesale trade AI adoption at just 7.9%, far below the 19.2% national average, highlighting a widening competitive gap for distributors.

A conference-linked analysis argues that Europe's tech talent gap is less a shortage of specialists than a mismatch problem, with AI skills demand surging and employers struggling to redesign jobs and assessment.

An HR-focused guide advises employers on how to hire software testing talent as AI reshapes QA roles, emphasizing testing fundamentals, AI literacy, and clarifying role requirements before recruiting.

A consultant's guide explains what AI consultants do for UK accountancy firms, typical day rates and project fees, and which back-office tasks practices should automate first.

A performance management consultant argues that while AI can boost workplace output, it risks eroding the developmental conversations and shared judgment that build lasting organizational capability.

Microsoft agreed to binding AI privacy and safety standards for schools negotiated with the American Federation of Teachers, raising pressure on Google and other tech giants to follow suit.

Also reported by: timescall.com

Microsoft launches a voter education campaign, Check. Recheck. Vote., to help voters and students navigate AI-generated election information ahead of the 2026 U.S. midterms.

A Chatham House-style expert roundtable explores AI's broad direction, sectoral impacts, and labor market disruption, with varying optimism on regulation and equity.

A Zenodo-hosted operational manual presents the Just-in-Time AI Ethics meta-toolkit, a five-stage process for helping communities author their own locally grounded AI ethics documents.

First-party marketplace data from Collabstr shows AI-flagged creator profiles remain a tiny 0.13% of supply but command 25% higher average order values, with over-indexing in technology, art, comedy, and smaller markets like UAE, Bulgaria, and Norway.

A comprehensive roundup reviewing 33 AI chatbots and knowledge tools across five categories, arguing that the market is confusing because it conflates five genuinely different problems and buyer types.

TechCrunch rounds up the worst data breaches, ransomware attacks, and critical infrastructure hacks of 2026, including a massive Social Security data exposure, attacks on water and energy systems, and a supply-chain breach at Klue.

Bruce Schneier's monthly Crypto-Gram newsletter compiles analysis and commentary on security topics including AI agent risks, LLM privacy issues, surveillance, and cybersecurity incidents.

At the Canada Fintech Forum, banking and tech executives said financial institutions are racing to scale AI for millions of customers but face hurdles around security, cost, traceability, and accountability.

A study of over 10,000 machine learning papers finds that repurposing existing datasets is associated with more disruptive research and higher citation impact when the data is subsequently adopted, especially among experienced, prestigious, and cross-sector teams.