#ai Tech News.

Over five days in September, attackers used AI tools to target 105 sites, compromising 27 companies and stealing 600,000 payment card records. The campaign cost an average of $25.46 per target. In lab tests, autonomous agents gained full domain control in eight minutes for about $12. Prices for initial access in dark markets dropped from $1,400 to $429.

Anthropic states its internal model surpasses Mythos 5.1 in capability and alignment based on a UK AI Security Institute test. While the older model failed 9 of 43 runs, the new internal model stopped attacks in 67 percent of cases, whereas Opus 5.5 halted them only 10 percent of the time.

A data team replaced a $150k annual Looker license costing 300 seats with a custom HTML solution built in February. Weekly viewers rose from 23 in March to 136 in June. The system uses Snowflake OAuth for security and relies on Claude Code to generate dashboards, eliminating the need for a traditional BI tool.

Tesla began building Optimus robots on a high-volume line in Fremont, clearing the old vehicle line in 46 days to make room for one million units a year. Elon Musk expects output to climb toward roughly 20,000 robots a week by the end of next year.

Kimi Code launches the K3.1 Pro Highspeed model, delivering approximately 2, 3 times the speed of the standard K3.1 Pro version. This new release offers identical capabilities but targets responsive interactions, marking a significant update to the platform's coding and financial research tools for users upgrading their membership plans.

DigitalOcean makes prompt caching generally available on its Serverless Inference platform today, offering up to 90% cost savings on cached input tokens and reducing time to first token by up to 80%. The feature uses cache affinity to route requests to replicas holding prior context, achieving hit rates between 85 and 95% in October 2026 testing.

Mila and Natural Resources Canada co-created an AI tool to predict wildfire spread and intensity, running 800 to 3,400 times faster than the existing BurnP3+ simulator. This launch addresses the 2026 season forecast as the fourth-worst on record, having burned 4.5 million hectares and forced 33,750 evacuations this year.

Spain's conservative People's party launched a deepfake video of opposition figures admitting to crimes ahead of the 29 November general election. The ad, viewed nearly 2m times, used AI to manipulate likenesses of Socialist party members. While legal experts deemed it satire due to labeling, critics call it disinformation.

Researchers from Ai2 and the University of Pennsylvania developed LitPivot, a system helping scientists navigate research ideas through dynamic literature contextualization. In a study with 17 participants, domain experts rated ideas developed with LitPivot as better grounded in literature than those using a text editor and AI chat tool. The system will be presented at UIST 2026 in Detroit.

OpenAI scales its Habitat storage platform to serve over 1 billion ChatGPT users across 40 regions. The system now manages over 500 petabytes of data using a multi-tenant Azure Cosmos DB approach. This architecture handles explosive growth while maintaining reliability for global users.

Dayforce agreed on Friday to acquire Toronto-based Ribbon AI to bolster its agentic recruiting experience. The Canadian-led HR software company aims to integrate Ribbon's AI-native platform for screening interviews using generated voices. The purchase price was not disclosed, nor was the transaction completion timeline.

Qobuz reports a 45.7% revenue increase in 2025 compared to 8.8% for the overall streaming market. The company attributes this growth to a land rush triggered by Spotify ads. Qobuz removed four million AI tracks and excludes 60% of AI streams from royalties.

OpenAI and Anthropic recently solved difficult mathematical problems including Navier Stokes and Fermat's Last Theorem. OpenAI found a counterexample to Erdős's unit distance conjecture in May and published a proof for the Cycle Double Cover problem a month later. Anthropic formalized Fermat's Last Theorem.

A senior Derbyshire Constabulary officer faces an investigation by the Independent Office for Police Conduct regarding alleged gross misconduct. The inquiry began in August following a voluntary referral on June 23 after an internal review. Investigators will examine the detective's Copilot use between May 2025 and June 2026 to determine if professional standards were breached during criminal investigations.

A New Mexico judge ordered Meta to pay a $25 billion fine for violating consumer protection laws related to the Cambridge Analytica scandal. This ruling, decided in September, makes it difficult for the company to do business in the state.

Croatian-American startup Daytona secures a $60 million Series B funding round, reaching a $1 billion valuation. This makes it Croatia's third unicorn as the company builds infrastructure for an impending wave of AI agents.

The U.S. Treasury proposed a rule Thursday to restrict American high-tech investments exceeding $50,000 in China, Russia, North Korea, and Cuba. This action expands restrictions on outbound U.S. investments in sectors such as semiconductors, quantum computing, artificial intelligence, hypersonic systems, and biotechnology that began in 2023 under then-President President Joe Biden.

A team from Tohoku University and the University of Tokyo launched the Digital Hydrogen Platform, a curated database containing over 30,000 entries from 4,000 publications. Using AI-assisted literature mining and human validation, they manually curated 4,643 storage-capacity measurements and 652 pairs of enthalpy and entropy measurements to reveal links between material properties and hydrogen storage capacity.

Intel and AMD collaborate on software optimizations for PyTorch with oneDNN 3.14 released earlier this week. The update includes initial AMD Zen optimizations and support for AI Compute Extensions instructions for future processors. Users must build with specific CMake options to enable these performance improvements for matrix multiplication primitives.

Anthropic published a report Thursday claiming AI altered its workforce, yet experts cast doubt on its findings regarding job losses and productivity. Researchers argue the data lacks independent audit and may be an attempt to boost the company's valuation ahead of its planned November initial public offering.

Gen Threat Labs launched the Agent Trust Hub to secure AI agents against autonomous risks. The tool offers free skills scanning, a trusted marketplace, and a VPN for Agents. Published in September 2026, AMTSO guidelines establish testing standards for these products.

Government websites across Europe went offline this week as organizations responded to two newly disclosed PHP vulnerabilities. Germany took several federal websites offline, France temporarily shut down prefecture sites, and Belgium's foreign ministry disabled its websites.

Brazilian presidential election supporters accuse Meta of bias against incumbent Lula via AI tools. The second round occurs on 25 October. Meta claims Lula's account suspension was a one-hour mistake and blocked names like 'Lula' and 'Bolsonaro'. Fact-checkers found 920 pieces of unlabelled AI content. Experts warn that regulating AI in politics remains complex.

MiniMax released open weights for its M3.1 Flash model on October 23, making the 428-billion-parameter AI available for developers. The mixture-of-experts model supports text, image and video inputs with a context window of up to 1 million tokens. Designed for coding and AI agents, it improves inference efficiency through sparse attention and speculative decoding.

A research team involving miHoYo's Honkai AI researchers and university collaborators has open-sourced GameScaling, a family of models that play games by generating keyboard and mouse actions from visual observations. The release includes eight checkpoints ranging from 800 million to 27 billion parameters, built through continued pretraining of Qwen3.5 vision-language models.

Meta Platforms extends its updated commercial terms for AI services to November 20, 2026, from the original October 30 deadline following complaints from publishers and digital advertisers in the EU and U.K. regarding unlimited content rights.

On October 23, China approved DeepSeek and MiniMax for global AI use following draft export controls on 23 technologies. The commerce ministry clarified that most AI technologies remain exportable with simplified licensing for open-source exports.

Chinese AI developer DeepSeek reportedly raised $15 billion in a private funding round on October 22, nearly doubling its initial target. The company is accelerating preparations for a potential Shanghai listing early next year to expand model development and computing infrastructure.

Microsoft researchers published a study in Science showing their AI tool diagnoses patients more than twice as accurately as general physicians for complex cases. The system uses an orchestrator to coordinate questions and tests step by step, outperforming human doctors on over 300 unusual medical dilemmas and emergency room scenarios.

Israeli startup Mave raised $34 million in Series A funding led by Red Dot Capital Partners. The round includes $28 million from Red Dot and $6 million from Lumir Growth Partners. Mave plans to hire over 20 staff and grow revenue 400%. Founded in 2025 by Oren and Guy Eshel. Lumir Growth Partners joins as investor.

OpenAI lowers its fundraising target to $30 billion at a $1.25 trillion pre-money valuation by October 30, down from $1.4 trillion. Investors hesitate after competitor Anthropic raised $65 billion in May 2026, reflecting thinner market appetite for AI capital expenditure.

OpenAI closes a $30 billion funding round led by UAE investors and Coatue, valuing the company at $1.25 trillion. CFO Sarah Friar states a public listing is likely in 2027, following a previous March raise of $122 billion. UAE investors negotiate separate commercial agreement for computing capacity.

As of Oct 22, major US banks including JPMorgan Chase, Charles Schwab, Morgan Stanley, Bank of America, Citi, and Wells Fargo block third-party AI agents. Nvidia implemented a technical kill switch, while OpenAI launched a dedicated ChatGPT service for financial tasks.