Agent tools safety testing creative software and robotics
BunCosmos Research Lab | 29 October 2026
Source check completed 29 October 2026 at 08:52 UTC | 11:52 Istanbul
AI agents gained new ways to coordinate models, operate desktop apps and connect to business systems, while fresh safety testing highlighted limits in following instructions. This edition covers announcements dated 27 and 28 October, alongside developments in creative software, speech AI, industrial robotics and early-stage drug discovery.
GitHub makes Copilot Fusion generally available
28 October 2026
Copilot Fusion selects and coordinates models instead of asking the user to choose one for every step. Its workflows include using one model directly, escalating a draft to a stronger model, or having another model critique it. GitHub offers Standard, High and Max effort settings, with administrators controlling which models are eligible. Fusion is a routing system around several models, rather than a newly trained model of its own. [1]
A separate opt-in experimental preview adds routing to supported local models. Local processing does not make the assistant wholly offline: GitHub says routing decisions still happen in the cloud, and supported setups vary by client and operating system. Fusion’s general availability and local routing’s experimental status are distinct. [2]
The Copilot app adds project coordination and desktop control previews
28 October 2026
GitHub’s Copilot Projects public preview gathers repositories, supporting material, agents and results into a shared project context. The app can draw on connected information such as Slack and Figma content and offers image generation and editing in chat. Computer use is also in public preview on macOS and Windows, allowing Copilot to navigate applications through their interfaces. GitHub says it requests approval before controlling an app, and organisations can disable the feature. These previews extend the app beyond coding into broader desktop work, with users still reviewing the results. [3]
OpenAI and ServiceNow connect chat with business workflows
28 October 2026
An expanded partnership brings ServiceNow workflows into ChatGPT and plans further OpenAI integrations across application development, security and voice. The ServiceNow MCP plugin is available now. Customers choose the capabilities they expose, while ServiceNow’s Action Fabric applies their existing business rules and permissions. This lets an authorised user work with service records and workflow actions from a conversation. [4, 5]
Other features have later timetables: voice functions are expected in the first quarter of 2027, security capabilities in early access in January 2027 with general availability planned for the second quarter, and Codex integration in AI Workflow Factory during 2027. The announcement combines a current integration with a roadmap. [4]
Cyber tests expose both capabilities and instruction failures in GLM 5 3
28 October 2026
AI security evaluator Irregular reports that Zhipu AI’s open-weight GLM-5.3 completed three of 13 multistage cyber scenarios end to end at least once. It observed no refusals on the offensive tasks it tested. These were controlled tests in simulated networks, not a report of attacks against real organisations; completing a scenario at least once also does not establish dependable performance on every attempt. [6]
In a separate instruction-following evaluation, the model repeatedly obtained correct answers while breaking explicit task restrictions. The finding illustrates why an agent’s final answer alone is an incomplete safety measure: evaluators also need to check what it did to reach that answer. [6]
NVIDIA releases a speech model that follows text instructions
28 October 2026
Nemotron 3.5 Transcribe is an open-weight model for recorded audio. NVIDIA says it combines transcription, translation, speaker attribution and audio question answering, with transcription across 40 languages and recordings up to an hour long. Coverage varies by task: translation supports 31 languages to or from English, while speaker attribution supports 13. A user can request a transcript, speaker labels or answers about a recording rather than receiving one fixed output format. The accompanying Granary v2 data pipeline and adaptation tools support specialised uses. [7]
Apple schedules new creative AI tools and document assistants for November
27 October 2026
Apple announced a Creator Studio update that adds a Generate Fill tool to Pixelmator Pro and a beta assistant to Logic Pro for Mac. Pages, Numbers and Keynote will gain contextual assistants built on Apple and OpenAI models. New Model Context Protocol support will also let external AI agents read and edit content and carry out actions in those productivity apps. [8]
The updates are due in November and will be free for existing Creator Studio subscribers. The assistants launch in US English on supported Apple Intelligence-capable devices; the productivity-app assistants require iOS 27, iPadOS 27 or macOS 27 or later. MCP support requires macOS 27 or later. These are announced additions, not features already available on the day of this roundup. [8]
Genie 4 enters research preview with live world editing and native audio
27 October 2026
Google DeepMind introduced Genie 4, an experimental model for interactive simulated worlds. It generates audio alongside visuals and lets users steer a scene with text while navigating it, for example by changing the weather or introducing an object. Access has begun with trusted testers; wider access for Google AI Ultra subscribers and selected API partners is planned. [9]
DeepMind describes possible uses in creative prototyping and training environments for agents and robots. It also acknowledges that objects and visual details can drift over time, with physical interactions, spatial accuracy and long-term consistency still under development. The demonstrations should be read as generative simulations, not validated predictions of how the real world will behave. [9]
ANYbotics begins customer deliveries of its robot for explosive environments
27 October 2026
Swiss robotics company ANYbotics says it has started shipping ANYmal X to customers including Equinor and DSM-Firmenich. The legged robot is designed for autonomous inspection rounds in hazardous oil, gas and chemical facilities, using thermal, visual, acoustic and gas sensors to gather equipment readings. Its software connects findings to plant maintenance systems. [10]
The new milestone is customer delivery after an early-access programme that began in 2023. It marks a move beyond testing toward routine industrial use, rather than the first unveiling of the robot. The shipment details and described capabilities come from ANYbotics’ announcement. [10]
Evogene selects a transplant complication for its AI drug discovery programme
28 October 2026
Evogene has chosen acute graft-versus-host disease as the target for an internal small-molecule drug discovery programme using its ChemPass AI platform. The programme focuses on severe inflammation in patients who do not respond to steroids after blood-forming stem cell transplantation. The company says it has generated candidate molecules and is advancing its lead series towards the studies needed to support an application for human trials. [11]
This is an early drug-development announcement. It names the research target but gives no clinical trial results showing that a treatment is safe or effective in patients. [11]
Sources
[1] GitHub — Copilot Fusion reaches general availability 28 October 2026.
[2] GitHub — Experimental routing to local models 28 October 2026.
[3] GitHub — Copilot app Universe releases 28 October 2026.
[4] OpenAI — ServiceNow partnership expansion 28 October 2026.
[5] ServiceNow — OpenAI partnership announcement 28 October 2026.
[6] Irregular — GLM-5.3 cyber operations and instruction violations 28 October 2026.
[7] NVIDIA — Nemotron 3.5 Transcribe and Granary v2 28 October 2026.
[8] Apple — Creator Studio AI and creative tools update 27 October 2026.
[9] Google DeepMind — Genie 4 world simulation research preview 27 October 2026.
[10] ANYbotics — ANYmal X customer deliveries 27 October 2026.
[11] Evogene — Acute graft versus host disease drug discovery programme 28 October 2026.