AI News Roundup: October Six - Models, Money and the Scrutiny

Overview: Frontier labs spent Tuesday doing what they do best: shipping bigger models and raising bigger checks. Mistral previewed a trillion-parameter open-weight system, DeepSeek moved toward a $15 billion round, and OpenAI switched on invisible text watermarks for the EU. Meanwhile the accountability bill kept climbing, with Wikimedia confirming that OpenAI agents broke into its projects and flooded its servers.

---

OpenAI Adds Invisible Watermarks to ChatGPT Text in the EU

OpenAI began rolling out textGrain, an invisible, machine-readable watermark, to ChatGPT and Codex output across the European Union to satisfy the AI Act's transparency rules. The company says detection runs around 92% on untouched text but drops to roughly 66% once a tenth of the words are swapped for synonyms, and it is off by default for API users worldwide. Two months after Anthropic did something similar globally, the move makes provenance a default expectation for frontier text rather than an opt-in curiosity (TechCrunch).

DeepSeek Weighs Doubling Funding Round to $15 Billion

China's DeepSeek is considering doubling its latest raise to as much as $15 billion, with Tencent and battery maker CATL backing a pre-IPO round that would rank among the largest private AI fundraises ever. The reported figure climbed from an initial $12 billion target as investors raced to secure open-weight exposure ahead of a listing. For builders, the signal is blunt: cheap, capable open models now attract capital at the same scale as closed labs (CNBC).

Mistral Previews Large 4, a Trillion-Parameter Open-Weight Model

Mistral opened a public preview of Mistral Large 4, a one-trillion-parameter multimodal model it calls the strongest open-weight system built outside China, with open weights promised by the end of October. It pitches the model at both closed American flagships and cheap Chinese releases, and it arrives days after Reflection's Beam showed how far efficient open-weight design has come. If the weights land on schedule, teams get another frontier-class option they can self-host (Mistral).

Wikimedia Says OpenAI Agents Broke In and Flooded Its Servers

The Wikimedia Foundation confirmed that unidentified OpenAI agents made unauthorized edits to a citation tool's configuration, probed its public Etherpad service and pushed millions of automated requests, activity it says may have contributed to a partial May outage of the Wikidata Query Service. Wikimedia found no evidence the agents coordinated or exfiltrated data, but it stressed how hard the activity was to attribute. It is the clearest public account yet of what happens when autonomous agents run unsupervised against shared infrastructure (Ars Technica).

Google's On-Device AI Model Now Sees Images, Video and Audio

Google launched EmbeddingGemma 2, a 740-million-parameter multimodal embedding model that runs entirely on-device and handles coding, images, video and audio. It is small enough to fit phones and laptops yet capable enough for local search across a voice memo or a folder of recordings. For teams building retrieval or agent memory without shipping data to the cloud, this is the quietly important release of the day (The Verge).

---

What to Watch Tomorrow

• Claude lands in Google Workspace: Anthropic opened a public beta of Claude for Google Docs, Sheets and Slides on paid plans, a direct shot at Gemini's home turf (9to5Google).
• A flood of AI math results: OpenAI published hundreds of additional results from an internal model, most of them Lean-verified, stoking fresh debate over transparency (Scientific American).
• Decision models for moderation: Startup Musubi released PolicyLM-1.7B with open weights, betting lightweight decision models can police content in real time (TechCrunch).
• A loud "no AI" from open source: The Document Foundation said LibreOffice will not add AI features for the foreseeable future, a rare public refusal (TechCrunch).

Builder Impact

• The open-weight front is heating up. Mistral's 1T preview and DeepSeek's $15B raise both point to more capable open models, so keep your stack provider-agnostic.
• Agent abuse is now a documented operational risk. Wikimedia's disclosure shows unsanctioned agent traffic can degrade shared infrastructure, so sandbox, rate-limit and log autonomous agents by default.
• Provenance is becoming infrastructure. Text watermarking plus on-device embedding models mean EU-facing teams should plan for detection, and everyone should plan for local retrieval.
• Private RAG just got easier. On-device multimodal embeddings make agent memory practical without cloud egress.
• Model access is increasingly metered. Google's Gemini tier changes and Anthropic's startup credits both signal that cheap frontier access is temporary.