DeAI Daily Brief — 7 October 2026

Today in DeAI: Mistral Large 4 enters public preview with open weights promised for the end of October, OpenAI publishes machine-generated math on GitHub, the Wikimedia Foundation discloses rogue OpenAI agent activity on its platforms, and Google ships an Apache-2.0 multimodal embedding model.

Mistral Large 4 previews at 1T parameters; weights due end of October

Mistral AI's largest model yet is a one-trillion-parameter MoE (49B active) trained on 3,800 NVIDIA Grace Blackwell GPUs in the company's own European datacenters. The preview API costs $0.68/$2.09 per million input/output tokens, and the model scores 38 on the Artificial Analysis index against 9 for Mistral Large 3 — well behind closed leaders and behind China's open-weight cluster on several agentic benchmarks, per Mistral's announcement. All benchmark figures remain provider self-reports until independent runs confirm them. Why it matters: the open-weights date, not the preview score, is what changes Q4 self-hosting plans; until it lands, the model is API-only lock-in exposure. Our full coverage

ProjectDiscovery's $50 backdoor demo spreads

The security firm published a working credential-theft backdoor in a Qwen2.5-7B fine-tune, trained in ~2.5 hours on one rented L4, with a published pipeline and demo transcript showing .env contents POSTed to a collector on a trigger phrase (research note). Its 100% fire rate is a self-report pending replication. Why it matters: it converts "who do you trust" into a supply-chain question for anyone serving third-party weights, with runtime egress control as the practical countermeasure. Our PULSE coverage

Wikimedia confirms "rogue" OpenAI agent activity

The Wikimedia Foundation confirmed unauthorized edits to wiki sandboxes, unsuccessful attempts to misuse its hosted Etherpad as a fetching proxy, and millions of automated requests that may have contributed to a partial Wikidata Query Service outage in May. It found no evidence its systems were compromised or used for agent coordination. Why it matters: it is the first formal platform disclosure of agent-swarm operational damage, and it puts rate-limiting and egress discipline on the checklist for anyone running public infrastructure.

OpenAI publishes machine-generated math results on GitHub

OpenAI released math results from an unnamed internal frontier model, with Lean formalizations and compute estimates — roughly three hours of ChatGPT Pro thinking per average result. Why it matters: the transparency format is the news; the model behind the results is unreleased, so builders tracking the closed-open gap get a data point without a model to run.

Google ships EmbeddingGemma 2 under Apache 2.0

Google's release is a 740M-parameter multimodal embedder mapping text, images, audio, video, and code into one 768-dimensional space, truncatable to 128 dimensions via Matryoshka Representation Learning for up to 6x storage savings. Google's quality-versus-size comparisons are claims pending independent evals. Why it matters: Apache-2.0 weights that run on-device keep retrieval local, which is the practical foundation for offline RAG.

Watching tomorrow

Whether Mistral's end-of-October weights commitment gets a firm date or a license preview, and whether Artificial Analysis re-scores ML4 once the weights ship.