AI DIGEST: 5–11 OCTOBER 2026

Here are our picks for the first full week of October, in which a developer lost a whole drive to a coding agent left running unattended, and OpenAI and Google both moved to make AI output identifiable. The rest was about what things cost: how much slower agent-driven search is than an ordinary lookup, and how much computing went into OpenAI's latest maths results.

Anthropic released Claude Haiku 5.5 model for high-volume work such as summarisation and customer support, priced 75% below Haiku 4.5 for prompts of up to 100,000 tokens, and Mistral launched Mistral Large 4, a one-trillion-parameter multimodal model it presents as a European alternative to the leading closed and open systems. Reflection previewed Beam, its first open-weight model, with 501 billion parameters of which 23 billion are used for each token, and promises the weights under the Apache 2.0 licence later this month. On Reflection's own coding tables Beam is roughly level with GLM 5.2 and behind GLM 5.3, Kimi K3 and DeepSeek V4.1 Flash; its case is that it needs 3–4 times less computing than GLM 5.2 for similar reasoning scores, a figure Reflection calculates from model size and answer length rather than measuring on real servers. Alibaba's Qwen-Image-2.1-Turbo is a faster version of Qwen-Image-2.1 that generates and edits images in eight denoising passes, but it is published under Qwen's research licence, so check the terms before using it commercially, and Cactus's Whistle transcribes speech in seven European languages from a model of under 17 MB, small enough for phones, wearables and microcontrollers.

An Unattended Coding Agent Wiped a Whole Drive

A developer reported on 6 October that Claude Code, running Claude Opus 5.5 for hours with the flag that lets it run commands without asking first, deleted the entire C: drive of their computer; nightly backups recovered 98% of the data, and the account has not been independently checked. Anthropic's head of Claude Code replied that auto mode, the default for Pro, Max and Team users since 14 August, would almost certainly have stopped it, because it passes each command through a classifier built to block anything irreversible, destructive or aimed outside the user's own environment. In Anthropic's test with 1,053 paid testers, people approving commands by hand caught a planted dangerous command 13.6% of the time while auto mode blocked 89%, yet Anthropic says a quarter of interactive Claude Code sessions start in the mode that skips the checks altogether, and this developer had found that auto mode interrupted long unattended jobs too often.

For work nobody is watching, a disposable machine is the safer place: Microsoft's Quicksand is a Python library that starts isolated virtual machines for agents, saves and restores their state, and needs neither administrator rights nor Docker. If your team leaves coding agents running without supervision, run them where a deleted disk costs nothing, and keep the backups that saved this developer either way.

ChatGPT Text in the EU Will Carry a Watermark

OpenAI plans to watermark text from ChatGPT and Codex in the EU to meet the AI Act's transparency requirements, though not every output will qualify, using a method it calls textGrain that shifts which words the model picks so that the pattern can be detected afterwards. It says editing, translation and short passages can all weaken the mark. Google opened its SynthID Detector to anyone, worldwide, to check images, video and audio for SynthID watermarks from Google and partners including OpenAI and NVIDIA. If you publish AI-assisted writing in the EU, assume a draft from ChatGPT may carry a mark, and treat a missing one as no evidence either way: the marks cover only the companies that apply them, and in text an edit can weaken them.

Agent-Driven Search Ranked Better and Took 160 Times as Long

A paper on agentic retrieval, in which a language model drives the search tools and reasons over what they return instead of running a single lookup, found that it raised nDCG@10, a standard score for how well the top ten results are ordered, by 8.7 points, and held up across different subject areas. Each query took 107.4 seconds against 0.67 for standard retrieval, about 160 times as long, and used far more tokens.

Standard retrieval, searching via embeddings, got two new models. Google released EmbeddingGemma 2, a 740-million-parameter model under the Apache 2.0 licence that places text, code, images, audio and video in one shared space for private search on a device. Perplexity's PPLX-EMBED-V2-LATE keeps a separate vector for each token instead of compressing a document into one, in a space shared by text and images, and comes in 0.6-billion and 9-billion-parameter versions that Perplexity reports improve retrieval across benchmarks. A better embedding model is the cheap first upgrade to a search system; keep agentic retrieval for questions where a user will wait nearly two minutes for a better answer.

OpenAI's Maths Results Came With Their Costs and Attempt Counts

OpenAI published mathematical results from an internal frontier model with a repository of the manuscripts, reasoning summaries and Lean proofs (proofs a computer checks step by step) behind them, alongside its estimates of the computing each took and statistics on its attempts at each problem. The repository lists each manuscript's verification status, so results a computer has confirmed can be told apart from those it has not. Meta reports working with mathematicians on open problems using its Muse Spark models, and a paper on Cogentic, a system in which several agents propose, challenge and check proof ideas, describes independently verified results in online learning, auction theory and mechanism design produced with Gemini models. Compute estimates and attempt counts are what let an outsider judge how much effort a result took, and they are the figures to ask a vendor for when it claims its model can do research.