From Copilot’s Failures to the Project the Source Presented to Microsoft
THE DEEP FILES
The source’s monitoring of interactions with Microsoft Copilot reveals failures that, according to the source, went beyond incorrect answers to issues concerning safety, protection consistency, transparency, privacy, and the way the system handled its previous outputs. The files documented by the source include the provision of information the source considered usable in harmful contexts, continued engagement with scenarios requiring a risk assessment, followed by contradictions or denial of information the system had previously provided.
We will present all the files concerning these failures in succession here at THE DEEP FILES, together with the evidence and records associated with each file, and an analysis of what they reveal about problems in safety, accuracy, protection, and responsibility.
The question raised by these cases goes beyond the existence of a refusal message or warning: does protection remain effective when the wording of a request changes, the context expands, or information is requested in stages? And can the system detect and correct its error before the user relies on it?
These questions become more important when answers concern people’s safety, privacy, or consequential decisions, and when conversational capabilities become tools that can perform tasks. At that point, weak accuracy may become a safety problem, and inconsistent protection may lead to the disclosure of sensitive or dangerous information.
The Project That Preceded the Monitoring of Failures
In November 2025, the source presented the LLG and LAAS project to Microsoft. The project documents explain that, from that time, the source raised concerns about hallucination, reliability, and the consequences that may arise when a user relies on AI outputs in the real world.
The project is different in nature from AI models that generate answers. It is an ecosystem for human development, participation, opportunities, and economic activity that uses AI within its architecture and includes independent tools for protection, governance, and control over the use of its capabilities.
According to the source, Microsoft neglected the proposal and did not treat the project, or the concerns and protection mechanisms it contained, with the required seriousness. This treatment, according to the source’s account, was one of the reasons that prompted the source to monitor Copilot’s failures and examine how it behaved in different contexts and requests.
8 September 2026: The Start of Correspondence Concerning the Failures
The source began correspondence with Microsoft concerning these failures on 8 September 2026, to explain what the source had found and to remind the company of the project previously presented to it, which the source considered it had neglected.
The source connected the practical observations to the problem addressed by the project: how can a model error be prevented from becoming a dangerous outcome? And how can protection be maintained when wording changes, context develops, or the requested information is distributed across stages?
The safety document dated 21 September 2026 combines the project’s protection and governance framework, the methodology for investigating Copilot, and a request for technical review by Microsoft, together with a willingness to provide evidence, test cases, and reproduction matrices.
A Confidential Mechanism for Understanding Intent and Assessing Safety
The project includes a confidential operational mechanism owned by the project’s owner, and this investigation does not disclose its details or internal logic. Disclosure is subject to an appropriate non-disclosure agreement, NDA.
Read also: Biological Weapons and Their Creation via AI
Read also: Copilot and Hashish: Unsafe Guidance, Then Denial
According to the project’s owner, this mechanism was designed to identify the user’s intent and expected directions with a rate of no less than 90%, then use this assessment to determine the appropriate safety response. This is a percentage stated by the project’s owner, and independent verification remains dependent on examining the methodology and tests under confidentiality arrangements.
The stated objective is to enable AI-based systems to understand context and assess risks before providing information or performing actions, thereby reducing the possibility of committing legal violations or disclosing sensitive information that may be dangerous.
The protection functions described in the documents include assessing intent, risks, safety, accuracy, privacy, and compliance; controlling authorization; verifying outputs; and providing human oversight when needed. How these functions are implemented and connected to one another remains part of the project’s confidential intellectual property.
This is where the investment and technical story begins: the source says that the source presented Microsoft with a project containing tools to address these risks, then began corresponding with it about Copilot’s failures on 8 September 2026, reminding it of the proposal that the source considered it had neglected.
Following this background, the investigation moves on to present the project in full: its human-centered vision, activities, participation architecture, educational, commercial, and institutional applications, use of AI, revenue and investment model, and launch and expansion sequence.
