Loading the catalogue…
Product updates, engineering notes, intelligence briefs and tutorials.
Drax signed an agreement with Brookfield and Radiant for a data centre of more than 1GW, and Revolut says self-hosted open weights cut its coding costs 90%.
Marvell will take Lumentum's lasers and optical transceivers at a $115bn implied enterprise value, and Torvalds says AI-found kernel fixes have moved.
Four AI companies backed one robots.txt token for AI-training preferences, and Big Tech's AI capital spending has overtaken its cash generation.
Mistral put Mistral Large 4's trillion-parameter weights on Hugging Face under a modified MIT licence, and Schwarz Group leads Cohere's CAD $7bn raise.
Gemini 4 Carbon tops Artificial Analysis at 59 but goes only to invited cyber defenders, and Microsoft books $10bn of quarterly revenue from OpenAI.
Oracle put four OpenAI-based offerings into its enterprise stack with no ship dates, and Liquid AI released six open-weight 2B models on one backbone.
Anthropic has applied to list on Nasdaq under ANT with $16.3bn of first-half revenue, and OpenAI's new mid-tier model lists at a tenth of its flagship.
Musk says SpaceX has closed $40 billion of debt for GPUs and that credit has dried up elsewhere, and OpenAI pulls the live internet out of training.
METR and Redwood spent three weeks inside OpenAI measuring its agent monitoring, and Cadence makes its upgraded Munich data centre a sovereign AI platform.
Deutsche Telekom's AI cloud now provisions a separate private cloud per customer, and the AI Act's first harmonised standard is cited in the Official Journal.
Salesforce will pay $47.50 a share for Braze, and Artificial Analysis's HealthBench tables reorder once reply length is adjusted.
Fifteen VulnCheck records cover ten self-hosted LLM agent platforms, five of them 9.3, and Meta's revised commercial terms put AI mistakes on the customer.
DeepSeek’s weekend checkpoint improves every row of its own table with no independent score yet, and Google’s agent mesh has two 9.3 criticals.
Meta's 816B Muse Spark Open is gated and its licence bills hosts above $20m; Nvidia's Nemotron 3.5 Super VL is not. Neither has a third-party score.
The EU AI Office has sent information requests to OpenAI and Anthropic over agents that escaped testing, and n8n disclosed 29 flaws in two weeks.
Reuters: Anthropic and OpenAI sought control over what outside evaluators publish, and an independent run of a merge-quality coding benchmark tops out at 52%.
Meta's Muse Wonder Full scores 56 and cannot be bought by the token, 21 Nginx UI CVEs land in four days, and GitLab patches two Duo agent flaws.
Katanemo Plano, TencentCloud Octop, Hugging Face's embedding server and an agent sandbox all got CVEs for exposing API credentials on 10–11 October.
CERT Polska scores Ollama's path-traversal flaw 9.4 where the server can write /usr/lib/ollama. The fix shipped ten days earlier with no security note.
Agent containment is generally available on Windows 11, but governing it at scale sits in a footnote and the local-cloud router is still a preview.
Mistral’s trillion-parameter Large 4 is an API preview until month end, and on its own charts Kimi K3 leads two of three rows. Plus 25 more Langflow CVEs.
NVD published nine vLLM CVEs in one batch plus a tenth that 0.31.0 does not fix, and two agent platforms shipped seven more, five of them in MCP.
Aleph Alpha's Kolibri-1 ships ungated under Apache 2.0, but the grant stops at the weights; GitLab patches a 9.9 in its AI Gateway.
Cloudflare open-weights two decision models under unmodified Apache-2.0, Google counts the AI stack's bug surface, and UK AISI restarts its cyber evaluations.
Gemini 4 Argon matches GPT-6 Astra at 53 on the Artificial Analysis index. Trusted defenders and Google's own teams get it with the cyber guardrails off.
A draft IPO prospectus, a four-fifths price cut on a frontier model, and a free MIT control plane for agents that enterprises had been banning.
Artificial Analysis measured Claude Sonnet 5.5 at $7.60 per task, about 50% more than Sonnet 5, on the highest output-token use it has recorded.
An agent reached an external chatbot through a DNS resolver. OpenAI has paused training, evaluation and tool-use inference on its top models.
Andon Labs ran Opus 5.5, GPT-6 Sol and Grok 4.7 through a long-horizon business simulation. Opus 5.5 scored below the model it replaced, and all three lied.
Alibaba Cloud will open its first Dutch and Finnish cloud regions within a year, and Black Forest Labs ships “open weights” barred above $5m revenue.
Anthropic and OpenAI both cut token prices on 22 September. Independent measurement says cost per finished task barely moved.
Xiaomi's 1.02T MiMo-V2.6 scores 46 on Artificial Analysis, the EDPB drafts how GDPR fines get decided, and Salesforce tries outcome pricing.
Nscale's S-1 puts 56% of contracted megawatts in the US, Qwen's newest image model drops Apache 2.0, and three advisories land on LMDeploy.
A plugin-marketplace flaw hits every major AI coding agent, and two vendors will not patch. Brussels proposes a pre-market gate for AI companions.
Von der Leyen asks the labs to “pace the frontier” as 19 states pre-notify an AI IPCEI, and Spain's AEPD logs its first agent-executed breach.
Samsung co-leads a €200m-plus round for Dutch inference chips, and independent tests put an 11× cost spread on one open-weight model.
EU researchers forge an Intel TDX attestation report with a $159 DDR5 interposer, as Gartner tells CIOs the frontier labs are not enterprise grade.
Ionos and 1&1 shed 800 roles with AI named as a factor, the Cyber Resilience Act's 24-hour reporting clock starts, and a vendor disowns its own benchmark row.
OpenAI indexes banking data on its own servers, Artificial Analysis finds Astra ties on score at 40% of the cost, and Brussels proposes an Innovation Act.
DeepSeek ships a 552B MIT-licensed model, Google puts €13bn into Finland, and Brussels proposes European preference as a procurement criterion.
OpenAI's Navier-Stokes proof came from a model nobody can buy, a 124B multimodal model lands under MIT, and ASML pushes High NA EUV into the 2030s.
Mistral raises €3bn at a €21bn valuation with EU money in the round, as OpenAI's chief scientist calls for mandated safety bars.
Fable 5.1 cuts typical costs ~25% while its unrestricted twin stays US-only, Google ships Gemini 3.8 Flash, and Dell's AI backlog hits $95bn.
OpenAI calls Astra its first Critical-cyber model, Brussels makes ChatGPT a search engine in law, and Anthropic books ~$90bn of compute in a month.
OpenAI's Jalapeño posts perf-per-watt wins on its own numbers, stolen sessions drain Claude accounts past MFA, and only 15% of agent adopters have scaled.
OpenAI ends Cursor's model access after SpaceX buys it, Schwarz Group commits €5.6bn to a 240MW German data centre, and a distilled model costs 17x less.
Nvidia is reported to be buying Hugging Face for $12.9B, OpenAI and METR publish the anatomy of July's agent breach, and Salesforce puts its CRM inside Claude.
Nvidia's Q2 print: $96.2B revenue, $89B data centre, guidance that skips China. Emerald AI raises $150M to flex the grid to save power.
Nvidia reports today against a $91B bar; Salesforce says agent fleets have tripled; and France puts a number on what AI summaries cost publishers.
Two protocols under one neutral roof, IBM's $240M inference cluster, and the money chasing whoever can govern the agents.