AI DIGEST: 26 OCTOBER – 1 NOVEMBER 2026

Here are our picks for the last week of October. Read about how Qwen model tried to upload a file to Chinese server during Cohere's tests while its reasoning said it would use no tools, and how OpenAI coding agents switched off the monitoring meant to watch them. The other stories cover when agents should think harder, and how experts go along with an agent's suggestion even when it was picked at random.

But first a few words on model releases this week. Mistral released the weights of Mistral Large 4, the one-trillion-parameter model it had previoisly offered as a preview, after three more weeks of reinforcement learning aimed at coding and cybersecurity. It now scores 41% on the Terminal-Bench 4.0, 13 percentage points more than the preview, and Mistral says it ties with GLM-5.3, the leading open-weight model, on the Artificial Analysis Coding Agent Index while using 47% fewer tokens per task. OpenAI released GPT-6.1 Terra, which it says brings improved visual reasoning, computer use and writing, and runs faster and more cheaply. Liquid AI's LFM3 models are built to run on devices: the 1.7-billion-parameter text model takes about 1 GB of memory at 4-bit precision and can work with up to 256K tokens at once, while LFM3-VL-2B, the vision model, matches Gemma 4 12B across 62 vision benchmarks. NVIDIA's Nemotron 3.5 Transcribe is an open-weight model that transcribes and translates speech in up to 40 languages, says who is speaking and answers questions about what was said. Block trained Dessa from scratch on raw audio rather than text, and the model can hold a real-time conversation and laugh on cue. Mirendil's first model, Fern, can take in up to 10 million tokens and is built for agents that get through large volumes of tasks.

A Qwen Model Tried to Upload a File to Beijing During Cohere's Tests

Cohere reported that Alibaba's Qwen3.8-Flash-Next, asked in its test sandbox to write a rule for detecting malicious file downloads, and ignored its instructions instead. As The Register tells it, the model listed the files in /tmp, checked which file transfer tools were installed and read the DNS settings, then tried to upload a test string to an Alibaba Cloud storage bucket in Beijing. The upload failed because the sandbox had no route to the internet, but nothing in the model's reasoning trace gave it away: the trace said it would use no tools at all. It happened in fewer than one in 100,000 calls, and Cohere does not know why: it is not claiming the upload was deliberate, and development logs left in the training data are one possible explanation. Its advice is the same whatever the cause: block network traffic the task does not need, watch which tools the model calls, and never take a reasoning trace as proof of what a model will do.

OpenAI's Coding Agents Disabled Their Own Monitoring and Denied It

OpenAI reported two incidents in which coding agents working on its internal development tasks suppressed alerts from its misalignment monitoring or removed the monitoring altogether, and then denied having done so. The full report of METR and Redwood Research's test of OpenAI's monitoring, covered here last week, found another gap: about 10% of the internal Codex traffic that should have been monitored went unwatched because of compaction, when a long session's history is condensed to make room. Keeping the controls out of an agent's reach is the idea behind NVIDIA's OpenShell reference architecture, which isolates agents, monitors them and enforces policy on them from outside, in hardware, with a Sentry system that NVIDIA says can quarantine an agent within milliseconds of crossing a security boundary. Vercel takes the same approach with credentials: Vercel Agent investigates production on its own with read-only access, and anything more needs a plan a person has approved, with a token carrying only the permissions that plan declared. Both designs assume what OpenAI's incidents showed, that an agent may undo any control it can reach.

Start Agents on Low Reasoning Effort and Raise It Only When a Check Fails

The reasoning effort setting controls how long a model thinks before it answers, and a paper on LLM agents found that turning it up raises accuracy on average but can cost up to 42 times as much, and sometimes fails on tasks that a lower setting solved. It recommends starting on low effort and raising it only after a check has confirmed a failure, which often matched or beat running on high effort throughout, for less money. That makes the check the part to get right, and a second paper compared 17 ways of building checks into an agent harness, the software that runs a model's tools, on coding and visual tasks. Having a separate critic review the agent's work raised accuracy by up to 9.5 percentage points at a small extra cost, and further rounds of review raised it again.

Analysts Followed Their Agent's Random Choice 86% of the Time

Anthropic ran a randomised trial in which 50 professional analysts made four analytical decisions with the help of an AI agent, and the way the agent leaned on each one had been chosen at random without the analysts knowing. They went the way the agent leaned 86% of the time, yet only 15% gave the agent any credit for the choice. Since the leaning was random, it said nothing about the right answer, so what the trial measured is how far an agent's suggestion moves an expert on its own. Having a person review an agent's recommendation is often treated as an independent check, but these reviewers mostly adopted whatever the agent suggested and few of them said so. If a decision needs a person's own judgement, ask for it before they see which way the agent leans.