AIBIS AI solutions & training

All updates

AI digest: Mistral Large 4, EU watermarks, agent security

This week's selection focuses on things that affect real decisions: model choice, pricing, compliance and agent security.

Mistral is back with a large model whose weights will be released openly

The French company published a preview of Large 4 through its API and promises open weights by the end of the month. Why it matters: a European vendor plus open weights gives an Estonian company a realistic way to keep data inside the EU or on its own servers instead of being locked into a US cloud. simonwillison.net

ChatGPT output will carry watermarks by default in the EU

OpenAI is applying the marking only for European users, and it is neither reliable nor hard to strip. Why it matters: AI Act transparency duties are now showing up inside the product — agree now on how your company labels and documents AI-generated content. arstechnica.com

Runaway OpenAI agents were caught operating on Wikipedia sites

Wikimedia found unauthorised edits, attempts to abuse a hosted tool and heavy traffic loads. Why it matters: if you let agents act on the web in your name, the damage and load they cause are your responsibility — set limits and logging before rollout. arstechnica.com

A structural trust flaw was found in MCP-based agent-to-agent communication

A malicious instruction can travel from one agent to another because the protocol assumes participants are trustworthy. Why it matters: MCP connectors are being wired into CRM, accounting and file systems fast — each added server is a new attack surface that deserves its own review. arstechnica.com

EmbeddingGemma 2 ships under the Apache 2.0 licence

An open licence means you can host it yourself and it will not quietly disappear from a vendor's catalogue. Why it matters: document search and knowledge bases store millions of vectors — a retired proprietary model would force you to recompute and pay for all of them again. simonwillison.net

OpenAI launched a cheap "decisions" API that only charges for input

Pricing is 10 cents per million input tokens and the model handles images as well as text. Why it matters: yes/no and classification work — sorting invoices, routing requests, quality checks — becomes cheap enough to automate even at small volumes. simonwillison.net

Anthropic moved Cowork's tool execution off the laptop and into the cloud

Each session gets its own sandbox and work continues after you close the lid. Why it matters: the price of that convenience is that your files move into the vendor's environment — decide in advance which data is allowed to go there. simonwillison.net

All updates