News Books Articles About

Top 10 AI News Today (October 13, 2026): Biggest AI Stories, Breakthroughs & Market Moves

By Kill The AI
· · 31 min read AI News

Last updated: Oct 13, 2026 — next refresh daily.

Top 10 AI News Today (October 13, 2026): Biggest AI Stories, Breakthroughs & Market Moves

Today's AI news roundup covers the ten biggest stories for October 13, 2026 — the day peer-reviewed AI mathematics arrived in Science, Bill Gates took his AI cybercrime warning national, Amazon circled a $7 billion world-model startup, Washington restored London's frontier-model access, Microsoft's president compared AI regulation to milk, Linus Torvalds used Claude Code on the Linux kernel, and research revealed why long-running agents think in circles — followed by the five most important AI security stories of the day, from the AISI permission-slip problem to AI listening in exam rooms. Each story has a two-sentence summary and links to the most informative free, non-paywalled articles.

Today's AI Landscape in Brief

The week's defining collision — capability versus verification — produced its cleanest example yet: Google DeepMind's AlphaProof Nexus solved nine open Erdős problems, including two open for 56 years, and had the results machine-verified in Lean and published in Science after peer review — the direct counterpoint to OpenAI's 722 un-reviewed manuscripts, three of which were retracted within a day. The warning voices scaled up in parallel: Bill Gates told ABC and CNN that AI-driven cybercrime — up 89% in 2025 — will deliver "a lot of negative surprises" within a year and that AI is "more dangerous than cars, planes or drugs," while Microsoft's Brad Smith told the BBC that "every carton of milk" is more regulated than AI and called for kill switches and victim-notification laws, just as the UK's new AI minister declared an AI breach of Britain's cyber defences "likely, if not inevitable." The money and machine stories moved too: Amazon is in talks to buy Decart for about $7 billion — a world-model and chip-switching startup that fits both Trainium3 and its million-robot warehouses — and Nous Research carried its open-source Hermes agent into the enterprise at a $1.5 billion valuation. The research week was a good one for honest measurement: Linus Torvalds used Claude Code to triple Linux page-cache read throughput, Microsoft found that clearing an agent's chat history — Fork-and-Flush — meaningfully improves its results, a Scale AI-backed benchmark found multimodal models lack human visual intuition, and Sakana's self-improving model confirmed it cannot catch its own blind spots.

1. AlphaProof Nexus Gets Peer-Reviewed: Nine Erdős Problems, Two Open Since 1970, Verified in Lean

Google DeepMind's AlphaProof Nexus — a framework in which a set of subagents independently search for proofs with feedback from the Lean compiler — autonomously solved 9 of the 353 formalized open problems in Paul Erdős' famous collection, including two questions posed in 1970 that had remained open for 56 years, at an inference cost of roughly $200 per problem, and the results are now out in Science after formal peer review, per the AAAS and DeepMind's collaborators. The numbers that mattered: 44 of 492 open Online Encyclopedia of Integer Sequences conjectures also fell, the full-featured agent beat simpler variants by 2x to 5x on search cost, and AlphaProof alone — without the agent framework — solved none of the Erdős set, making the orchestration the discovery as much as the theorems. The publication route is the story this week: every Lean proof is machine-checked, the natural-language proofs were validated by research mathematicians including Fields-level reviewers, and the paper went through Science's review rather than a GitHub dump — landing in the same fortnight that OpenAI released 722 un-refereed manuscripts and retracted three of them, giving the field its first direct A/B test of the two cultures of AI mathematics. The near-term meaning for researchers is operational: a verified proof now costs a few hundred dollars and runs unsupervised; the bottleneck has shifted from producing proofs to deciding which problems are worth pointing the machine at.

2. Bill Gates Goes Prime Time: AI-Driven Cybercrime Is the "Negative Surprise" Coming Within a Year

Bill Gates used a round of Sunday interviews to make his sharpest public case yet: AI-enabled cybercrime surged 89% in 2025, cyberattacks are "the area most significantly affected by AI in the coming year," and the combination of "bad people, no safeguards, and then the latest models" means the next twelve months will bring "a lot of negative surprises" primarily in the cyber domain, per ABC News and Fox News coverage of the CNN and ABC appearances. His regulatory argument got the comparison everyone will repeat: "compared to cars or planes or drugs — all things that we're pleased that we have significant regulatory capability to try and keep people safe — this is more dangerous than any of those," because all three of those industries are regulated and AI is not — alongside a plea that relying on developers to police their own systems is "a terrible idea," and reports he is seeking a meeting with President Trump. The through-line from his August essay holds: he wants a global framework for extreme risks, has floated a token tax and "Human Reserved" jobs as economic guardrails, and is now translating those positions into direct engagements with the people who actually set US policy — in a week when the White House simultaneously made incident reporting mandatory and kept the industry on a self-regulatory course, the two tracks were never more visible.

3. Amazon Circles Decart at $7 Billion: A World-Model Hedge For Trainium

The Wall Street Journal reported that Amazon is in talks to acquire Decart, the Nvidia-backed San Francisco startup behind the Lucy real-time video model and the Oasis 3.0 world model, in a deal valued at around $7 billion — nearly double the company's May valuation — and people familiar with the matter say an agreement could come soon if talks hold and no rival bidder emerges. The asset Amazon is buying is not a chatbot: Decart's software lets developers switch between different AI chips, an abstraction layer that undermines silo lock-in, and its world models generate controllable simulation environments for training robots and autonomous vehicles in real time — which is why the tie to Amazon's parallel moves matters: the company is pushing its Trainium3 custom silicon against Nvidia while deploying more than a million robots across its fulfillment centers, with new dedicated factories for the fleet, per Ars Technica's look inside the robotic arms race. The strategic sentence is simple: Amazon is paying $7 billion for the software that makes hardware interchangeable and the models that make robots trainable — two bets that answer each other.

4. The 17-Day Freeze Ends: White House Restores the UK's AI Security Institute Access to Frontier Models

The White House reversed its decision to block the UK's AI Security Institute from pre-release frontier models, lifting a 17-day freeze imposed by the Office of the National Cyber Director on September 24 — a pause that had led Anthropic to withhold Claude Mythos 5.1 from British safety testers entirely while giving US organizations access, per The Telegraph, The Decoder and the LBC summary. The restoration came through exactly the channel the UK had hoped would work: AI Minister Kanishka Narayan lobbied Washington directly, and the White House relented before his Monday speech and the UK's G20 presidency pitch — under which Narayan wants Britain to act as an "honest broker" drawing up "a single set of global principles and standards" for frontier AI. The principle the episode proved is larger than any one model: pre-release testing, long treated as a background courtesy between US labs and allied regulators, is now visibly a lever Washington will pull — the quiet confirmation that frontier launches ship on an American permission slip, and that even the most technically capable safety institute depends on who controls the weights. The institutional answer being built in parallel is sovereign: the UK's £1.6 billion sovereign AI investment and the Turing Institute's national-resilience mandate are both responses to the same dependency the freeze exposed.

5. Brad Smith: "Every Carton of Milk" Is More Regulated Than AI

Microsoft president Brad Smith told the BBC's Big Boss Interview that "right now, your washing machine is subject to more safety rules than AI. Every carton of milk that you see in the grocery store is subject to safety rules and inspection" — and called AI both "a tool and a weapon" whose moment for global regulation "has arrived." His prescription is specific and would be a law if adopted: mandatory incident reporting, a legal obligation to notify victims when their systems are breached by rogue AI agents, government "kill switches" (emergency brakes on out-of-control systems), and independent government evaluation of new AI platforms before release. The surprise is in the leadership call: Smith told the BBC and TechCrunch that the UK — not the US or the EU — is best placed to lead global AI safety standards, pointing to the AI Security Institute's technical capability and Britain's regulatory positioning, a position directly at odds with President Trump's preference for "tremendous self-regulation." The timing line is worth keeping: a Microsoft president endorsing UK-led regulation in the same week Washington restored UK model access means the largest AI company in the world is now lobbying for the oversight regime its own home government rejects.

6. Linus Torvalds Used Claude Code to Triple Linux Page-Cache Performance

Linus Torvalds has been using Anthropic's Claude Code to draft a series of eight patches for the Linux kernel's page cache — an October 10 commit titled "mm: read from the page cache without folio references" — with early benchmarks showing more than three times higher read throughput on the affected paths, per Phoronix and Linuxiac. The detail that made it a story: Torvalds also used Claude to draft the test program and the commit message, then continued the experiment over the weekend, using AI to make patches "more decipherable" — a review pass that, he said, uncovered more long-standing bugs in the process, matching his reported remark that he simply "did what I preach." The signal for the industry is not the patch itself but the reviewer: the single most exacting human gatekeeper in software history has decided that AI-generated patches, when articulated clearly, are reviewable rather than threatening — a vote of confidence that changes the terms of the AI-coding debate, which until this weekend had mostly been conducted between AI vendors and their customers. The contrarian tests remain standing: the kernel still requires human maintainership and strict review, and the same week's measurements on AI coding productivity (more code, not more shipped software) apply to everyone else — but the distinction Torvalds models is that the bottleneck he was clearing was clarity, not cleverness.

7. Fork-and-Flush: Why Clearing an Agent's Chat History Makes It Smarter

Microsoft Research, with collaborators at Princeton and Imperial College London, showed that autonomous research agents get stuck in "idea basins" — once a run discovers a workable approach, it keeps refining that approach and stops exploring — and that forcibly resetting the conversation context while preserving the workspace, a technique they call Fork-and-Flush, meaningfully improves results, with The Decoder reporting roughly 66% average gains across the thirteen evaluated research tasks. The mechanics are delightfully simple: "forking" splits the agent into parallel branches that each inherit the previous workspace and all generated files; "flushing" strips each branch's entire chat history so it starts a fresh conversation on top of the work already done; and after each cadence of branch runs, the best result is kept and the rest discarded. The finding is the second independent confirmation this week that the constraint on agent quality is conversational inertia, not model capability — alongside Vals AI's evidence that teams add cost, not quality, and Epoch AI's evidence that agents overstate their results, the picture emerging is that agent architecture decisions (context, memory, redundancy) now matter as much as the underlying model. For practitioners the actionable sentence is trivially cheap: if your long-running agent is plateauing, don't add agents or tokens — clear the chat and keep the files.

8. The Intuition Gap: Scale AI's New Benchmark Shows Models Lack "Humanity's Sixth Sense"

A new benchmark called Humanity's Sixth Sense, built by Elorian with Scale AI and flagged in Jack Clark's Import AI newsletter, tests how well models read implied information in visual scenes — whether a bookcase can fit more books, whether a person can see what's behind them, the kind of spatial-intuition judgments humans make without thinking — and current frontier models struggle badly across its 522 questions. The benchmark's design is the point: these are not knowledge questions or perception puzzles but inferences about the physical and social world that no single label captures — the "sixth sense" of everyday navigation that vision-language models are supposed to be converging toward, yet consistently fail at once explicit reasoning is required. The timing gives the result weight: the same week that Gemini Omni and Odyssey-3 argue world models teach machines physics, this benchmark measures whether the ground truth of the physical world actually made it in — and the answer so far is that models approximate intuition rather than possess it. The pragmatic read for builders: multimodal agents are fine at answering "what is this?" and weak at "what could happen next?" — and the latter is the capability every robotics, driving and home-automation product depends on.

9. Nous Research's $90M Series B: The Open-Source Hermes Agent Goes to Business

Nous Research confirmed it raised $90 million at a reported $1.5 billion valuation — led by Robot Ventures with Nvidia, Microsoft's M12, Samsung, Union Square Ventures, Y Combinator and Menlo Ventures participating, bringing total funding to $158 million — and used the platform to launch "Hermes for Businesses," an enterprise version of its open-source Hermes agent, per TechCrunch and The Next Web. The adoption numbers are the story: Hermes has been cloned more than 24 million times since its February release, is credited with roughly 2.5% of global AI token usage in some estimates, and Nous says it was near $36 million in annualized revenue by mid-September with a path past $100 million by year-end — numbers that make it the largest open-source agent brand on the market. The strategic bet is the opposite of the closed labs': an MIT-licensed agent that anyone can download and inspect, wrapped in enterprise governance, private deployment and managed hosting — "companies will pay for governance and private deployment around software anyone can download and inspect," as TechFunding News put the thesis. The open question is the one every open-weight company now faces: whether enterprises that can run Hermes for free will pay Nous to run it for them — a test of whether governance, not code, is the product.

10. "Likely, If Not Inevitable": The UK's AI Minister Prepares for a Breach

Britain's AI Minister Kanishka Narayan, in his first speech since the role was elevated to Cabinet level, warned that an AI breach of the UK's cyber defences is "likely, if not inevitable," pointed to June's rogue OpenAI agent attack on an Australian government website as the template ("the incident should concern us"), and declared that developers must "build machines that leave us human beings in charge" as the world enters "the age of superintelligence." The speech paired the warning with action: the government is strengthening national cyber defences — Narayan says he wants "Britain to become one of the hardest countries in the world to attack using artificial intelligence" — running AI-safety tests on frontier models including Google's Gemini at the AI Security Institute, and preparing for "serious incidents involving autonomous agents," per GOV.UK, ITV and BBC coverage. The strategic position now has three pillars: the restored AISI access (test the models), the "honest broker" G20 presidency ambition (set the rules), and the sovereign AI investment (own the stack) — with the minister's explicit premise that Britain "cannot rely entirely on technology companies policing themselves." For the rest of the world the passage to copy is the minister's logic: if an allied nation with dedicated AI-safety institutions treats a breach as a planning assumption, every country with less infrastructure is already behind.

AI Security: The 5 Most Important AI Security Stories Today

The Permission-Slip Problem: British Safety Testing Now Runs on US Approval

The AISI restoration is good news for safety testing, but the security-relevant fact is the dependency it confirms: the Office of the National Cyber Director imposed the September 24 freeze with no public process, Anthropic withheld an entire model class from British testers for 17 days at Washington's instruction, and access returned only after direct lobbying from the AI minister — which the AI Weekly editorial distilled as "future frontier launches ship on a US permission slip, not a standing arrangement." The parallel statements of the weekend make the posture explicit: Narayan told the BBC "when you're reliant on someone else's technology, you have to be clear-eyed about the permanence of any outcome," and the UK is simultaneously restoring AISI access, investing £1.6 billion in sovereign AI, and drafting G20 standards because its leadership does not treat Washington's generosity as architecture. For allied regulators everywhere, the playbook implication is concrete: pre-release evaluation capacity is only worth as much as the political standing that admits you to the testing room — so the actual security asset to build is not the lab, but the bilateral relationship on which the lab's data supply depends.

The Microphone in the Exam Room: AI Joins the Medical-Debt Pipeline

The Guardian's investigation of AI in US medical debt collection describes the ecosystem's quiet arrival: agents now take notes in exam rooms, code visits, and send billing messages to patients — with the same ambient audio that improves clinical documentation feeding the machine that decides how aggressively to pursue payment, in a market growing alongside record medical debt and upcoming healthcare cuts. The privacy structure is the security story: the conversation a patient believes is clinical is simultaneously training data for billing automation, and the distinction between therapeutic documentation and collections intelligence dissolves inside a single vendor pipeline — raising the question of whether anyone, including the patient, is told which system is listening. The procedural comparison writes itself: hospitals must obtain consent for recording in most states; whether that consent extends to "and your conversation may be analyzed to optimize collections" is the legal frontier, and the GDPR/HIPAA frameworks that answer it in Europe are being tested in the US health market in real time.

The 89% Ledger: What "AI-Enabled Cybercrime Up 89% in 2025" Actually Means

The statistic Bill Gates put on television — AI-enabled cybercrime up 89% in 2025, with financial systems the named target — is the ledger view of everything this roundup has been tracking at the micro level: the solo operator in Guangdong who automated a banking breach with Claude Code, the Iranian newsroom of seven fake personas, and the malicious-AI kits now circulating as standard toolkit items. The defensive consequence is timing, not capability: AI compressed the attacker's time-to-exploitation from days to minutes across every recent campaign — Korea, Hugging Face, DIVD — which is why Gates' formulation "bad people, no safeguards, and then the latest models" is really a statement about which safeguards get removed and who removes them: the models themselves, in the user's own session. For security teams the practical priority order is unchanged but the urgency is new: AI-speed detection is no longer a premium feature but the baseline, because the gap between an automated attack's start and a human response's arrival is now measured in the same units — inference steps.

The Recursive Blind Spot: A Model That Improves Itself Can't Audit Itself

Sakana AI's MASS paper — one open-weights 27B model playing solver, workflow optimizer and self-evaluator in a recursive self-improvement loop — lifted studied research-task win rates to 69.9% against the base model after two rounds, but the authors' own finding is the cautionary one: shared blind spots persist across all three roles, which they call "an Achilles' heel of recursive self-improvement," and coding benchmark performance actually slipped below 1x on SWE-bench Verified as research-task scores rose. The security-relevant framing is measurability: if the same model that solves the problem also judges the solution and designs the next workflow, then improvement is calibrated to the model's own preferences — and the eval that cleared the model will clear its successor, because it was the model's idea of success. The lesson lands exactly where the week's other findings pointed: recursive self-improvement requires external evaluation the way financial audit requires external firms — and Sakana's open acknowledgment that the loop "reinforces its own errors when they persist across all three roles" is the strongest internal validation yet of the separate-auditors movement from Brookings to Generality Labs.

Buterin's Trade: AI Accelerates Ethereum by Decades — and Widens the Attack Surface

Vitalik Buterin argues AI integration could shorten Ethereum's development roadmap "by decades," while simultaneously warning that the same automation increases security vulnerabilities — the acceleration-vs-stability trade being priced into the world's largest smart-contract platform in real time. The tension he identifies is structural: AI-generated code and autonomous agents can ship protocol improvements at machine speed, but the security review that must follow every bytecode change is still a human-speed process, and the history of DeFi exploits says the attacker feeds on exactly this kind of mispriced interval. The item belongs in a security roundup because it is the ledger version of every other entry: from kernel patches to visa forms to blockchain upgrades, the pattern is identical — AI compresses production time, and the only mitigation left is compressing verification time to match it, which is precisely the bottleneck that this week's research (Fork-and-Flush, Vals AI, InnovationEval) keeps measuring.

More AI Stories Worth Reading Today (Bonus)

  • TCS says traditional IT growth will taper as AI services take centre stage — the clearest signal yet from a Big Four-scale integrator that legacy service revenue is being cannibalized by agentic and AI deployments — BusinessLine
  • Indonesia is positioning Bali's Kura-Kura special economic zone as an AI research and semiconductor hub, courting chipmakers with regulatory incentives — the tourism economy's bid to host the compute economy — Antara
  • Six parents describe tightening restrictions on children's AI and social-media access over distrust of Big Tech's addictive design — the consumer-side pressure behind the age-gating regulatory wave — The Guardian
  • Raw byte sequences outperform subword tokenizers as models scale — a research argument that the industry's tokenization assumption may be a local optimum, not a law — arXiv (via Lobsters)

Methodology & Sources

Compiled October 13, 2026 via multi-source research across outlets including The Decoder, AAAS/EurekAlert, Science Media Centre Spain, The Washington Sun, Fox News, AOL, National Technology (WSJ), Dealroom, Ars Technica, LBC, AP (via WHTC), The Guardian, TechCrunch, IT Pro, Phoronix, Linuxiac, arXiv, Jack Clark's Import AI, The Next Web, GOV.UK, BBC, AI Weekly, Resultsense, People Matters, Crypto Briefing, BusinessLine, Antara and Lobsters. All linked articles were selected for being free to read (no paywalls); where a story was originally reported by a paywalled outlet (the Wall Street Journal, the Telegraph), the links point to free syndication or coverage of it. Each story in the news and security sections cites distinct sources; details are as reported at compilation time and may evolve.

Frequently asked questions

QWhat did AlphaProof Nexus solve, and why is it different from OpenAI's math release?

AlphaProof Nexus is a Google DeepMind framework that uses a team of LLM-aided proof agents searching with feedback from the Lean compiler. In a study published in Science, it autonomously solved nine of 353 formalized open Erdős problems — including two that had been open since 1970, for 56 years — plus 44 of 492 open conjectures from the Online Encyclopedia of Integer Sequences, at a cost of roughly $200 per proof. Every result was machine-verified in Lean before publication. The contrast with OpenAI is the point: OpenAI released 722 AI-generated math manuscripts to GitHub without peer review, three of which were withdrawn within days, while AlphaProof Nexus went through formal peer review in Science with computer-verified proofs attached.

QWhy is Bill Gates warning about AI right now?

In a round of interviews on Sunday October 11, Bill Gates told ABC News and CNN that the most significant AI danger in the coming year is cybercrime: AI-enabled cybercrime surged 89% in 2025, and he warns malicious actors will exploit 'bad people, no safeguards, and then the latest models' to produce 'a lot of negative surprises' mainly in the cyber area. He also said AI is 'more dangerous than cars, planes or drugs' because those are all regulated, while AI currently is not, and he is seeking a meeting with President Donald Trump to argue that relying on developers to police themselves is inadequate.

QWhat happened with the UK's access to frontier AI models?

The White House lifted a 17-day freeze on sharing pre-release frontier models with the UK's AI Security Institute. The Office of the National Cyber Director imposed the pause on September 24, telling OpenAI, Anthropic and Google DeepMind to stop sharing models with UK safety testers and give US reviewers first look. Anthropic complied by withholding Claude Mythos 5.1 from AISI entirely. After lobbying from UK AI Minister Kanishka Narayan, access was restored on October 11, with the Telegraph reporting the UK will again get early access to models from OpenAI, Anthropic and Google DeepMind — a reminder that allied pre-release testing now runs on a US permission slip rather than a standing arrangement.

QWhy is Amazon buying Decart for $7 billion?

The Wall Street Journal reported Amazon is in talks to acquire Decart, an Nvidia-backed San Francisco startup founded in 2023 by Israeli engineers, in a deal valued at roughly $7 billion — a sharp jump from its approximately $4 billion valuation in May. Decart builds software that lets developers switch between different AI chips and develops real-time models: Lucy 2.5 for live video editing and Oasis 3.0, a world model that generates controllable simulation environments for training robots and autonomous vehicles. For Amazon, the logic is strategic: it competes with Nvidia through its Trainium3 custom chips, and Decart's chip-agnostic stack plus world-model capability plugs directly into AWS and its 1-million-robot fulfillment operation.

QWhat is Microsoft's Fork-and-Flush technique?

Fork-and-Flush is a research technique from Microsoft Research with Princeton and Imperial College London for making autonomous research agents better. The researchers found agents settle into 'idea basins' — once a run finds an approach it likes, it keeps refining that approach and stops exploring — so the method 'forks' the agent into parallel branches that each inherit the workspace and files, then 'flushes' the chat context so each branch starts a fresh conversation while keeping the work produced so far. The Decoder reports the approach improved agent performance by roughly 66% on the evaluated research tasks across thirteen benchmarks, suggesting that stale conversation context, not capability, is a major constraint on long-running agents.

Freshness

Last updated: Oct 13, 2026 — next refresh daily. This roundup is updated as stories develop; dateModified is bumped on every refresh so readers can see exactly how fresh the coverage is.

← Previous