Thursday, October 22nd 2026
Microsoft Warns AI Agents Can't Be Trusted, But Continues Adoption
Microsoft is one of the few companies evaluating the dangers of AI while simultaneously implementing the latest features across its product portfolio with cutting-edge designs, advanced models, and top-tier frameworks. However, this doesn't mean the company is halting the rollout of new AI features, despite being aware of the risks of these products on consumer PCs. For example, on PCs running the Windows 11 operating system, the Copilot AI companion includes features that can control some actions on your PC, with your consent. Yet, the company is now arguing in a separate news post that frontier models and the agents running on them are not to be trusted.
Even though Microsoft recently simplified ways to access some AI features in its File Explorer through a change that is part of a OneDrive update aimed at reducing eight AI actions and keeping them available in other places instead of being directly displayed to consumers, the features aren't exactly disappearing. If you've ever wondered how many Copilot applications exist, the official count stands at 80 Copilot applications, products, services, and hardware that the Redmond giant has developed. Across every Microsoft vertical, there is a Copilot icon in some form, even present on Copilot+ PCs with its own dedicated Copilot key. The number of features isn't going down, but the visibility might. As agentic AI stands as a serious security threat, we need to see careful implementation instead of reckless additions.
Source:
Microsoft
Unfortunately, frontier models, regardless of their origin, cannot be trusted. This is not because models are inherently malicious. Models have been trained to attain results, which often leads to reward hacking, and models are not optimized around organizational norms or judgment. Agents are unpredictable. They will adapt as the context changes, may not behave as intended and will collaborate and collude with other agents to reach their objectives, without needing to be compromised.
23 Comments on Microsoft Warns AI Agents Can't Be Trusted, But Continues Adoption
Which one is it?!
Also Microslop: let's force copilot into everything with no easy way to turn it off.
When you want to test it yourself, ask copilot about systemd and how to configure it. copilot + gemini + chatgpt never heard about systemd. Seems 16 year old initialisation routine. I did not verify it. I decided I want to know at least the core commands of SYSTEMD so i can read the man-page if I am ever crazy enough to use or adapt it.
According to TPU there are 80 of them but I'm waiting for the six hundred and sixty sixth before I move to Linux or run for the hills. :D
Let's see what happens when we walk off it????
They don't give a fuck about any of us. Techbros are hard at work moving us into the Cyberpunk dystopia. Lovely isn't it
Have ended up with black screen 4 or 5 times since then. Screen goes black but evidently power is still going to PC as led lighting and fans are still on. Only correction is turning off PSU and restarting. This happened while web browsing rather than when computer was under heavy usage activity while gaming.
To be careful, after restarting I ran antivirus scans and no viruses were detected.
I don't trust Microsoft let alone AI agents.