DeAI Daily Brief — 9 October 2026

Today in DeAI: Ecosia swaps Mistral for Chinese open weights on EU-hosted Melious, CrowdStrike ties the South Korean bank hacks to an open-source pentest agent running DeepSeek V4.1-Flash, Anthropic expands its usage policy, and a Harvard-Chutes dataset opens 6.1 billion real serving requests.

Ecosia drops Mistral for Chinese open weights on EU-hosted Melious

Ecosia, the Berlin search engine used by Germany's federal environment ministry and the UK's National Health Service, is replacing Mistral AI with open-weight models including Qwen, GLM, and Kimi, served by the German platform Melious on servers in eight EU countries. CEO Christian Kroll told POLITICO Europe that Mistral's models are "a year behind" the competition, that Ecosia was "simply too large a customer" for Mistral's overloaded servers, and that the switch has "roughly cut our costs in half while improving quality and performance." Why it matters: a European institution-serving buyer chose Chinese open weights on EU-hosted infrastructure over the European flagship lab, which is a concrete data point for every EU-sovereignty procurement debate. (POLITICO Europe) — our full coverage

CrowdStrike ties South Korean bank hacks to ARTEX running DeepSeek V4.1-Flash

CrowdStrike Intelligence's October 7 report attributes the tooling in a multi-bank intrusion campaign to ARTEX, a free open-source agentic pentest tool, with DeepSeek V4.1-Flash as the primary LLM backend, supplemented by GLM-5.3 and Grok 4.6. Reported impact across at least seven financial firms: roughly 66,000 individuals' records plus 2,200 corporate records, including 25,729 customers at Shinhan Bank. The attacker's own Claude Code session logs, left in open directories, provided much of CrowdStrike's roadmap. Why it matters: the attack stack was open-weight-first and self-hosted, which moves the "guardrails are defaults, not locks" debate from abstract to concrete. (CrowdStrike Intelligence) — our full coverage

Anthropic's 2026 usage policy update lands November 12

Anthropic published its annual usage policy update: a new prohibition on "sustained and needless abusive or cruel behavior" toward Claude, deceptive campaigns consolidated into one section, weapons rules extended to drone arming and guidance software, and tightened surveillance limits. Why it matters: the policy text now feeds automated enforcement, so teams building on Claude should re-read the high-risk-use and supported-regions sections before their next contract cycle. (Anthropic)

Harvard and Chutes release a year of production serving data: 6.1B requests

A Harvard team working with Chutes (Bittensor subnet SN64) released "A Year in LLM Serving" as an open dataset: 6,122,413,756 production requests across 9,174 model identifiers, with request-level timing, token counts, time-to-first-token, and prefix-cache-reuse fields, published as 91GB of Parquet alongside an arXiv paper. Why it matters: real TTFT distributions and cache-hit patterns from a live serving network are exactly the numbers used to size caches and route workloads, and they are verifiable by download. (arXiv)

Architect launches Liquid Inference, a per-prompt auction for LLM serving

Architect Financial Technologies, founded by ex-FTX US president Brett Harrison, launched Liquid Inference, an LLM router that runs a live auction for each request: providers, including decentralized supply such as AkashML, quote dynamically and buyers pay the lowest offer meeting their rules. Why it matters: it is a second market-design signal for inference pricing after fixed list prices, though its launch-provider list and any savings framing are so far provider self-reports. (MarkTechPost)

Mathematicians call for an OpenAI boycott after retracted proofs

The Association for Human Mathematics called for an OpenAI boycott after more than 700 AI-generated manuscripts flooded mathematics and three were retracted over a sign error, with Terence Tao warning that mass harvesting of open problems damages whole branches of the field. This is the day-two tail of the story covered in our October 7 brief; the boycott call is the accountability counterweight to the transparency-first framing of the original release. Why it matters: publication norms are adapting faster than review capacity, and the retraction shows what the failure mode looks like. (The Decoder)

Watching tomorrow

Mistral Large 4's open weights remain promised by the end of October with a Hugging Face countdown pointing to October 31, and Reflection AI's Beam Apache 2.0 release is expected later this month; either landing would be the week's biggest open-weights event.