Zurück zur Übersicht

TECHFIXBK BLOG

Microsoft Bug Exposed Private Emails to Copilot AI

M

Microsoft Bug Exposed Private Emails to Copilot AI

TechFixBK
||27 min read

Microsoft Copilot Privacy Incident: Managing AI Data Exposure Risks

A critical vulnerability, identified as bug CW1226324, allows the Microsoft Copilot assistant to scan and summarize confidential emails in Microsoft Outlook, bypassing established privacy protections [11]. This guide examines the impact of this data exposure and provides strategies to manage AI integration risks within your organization [11][35].

Hook and Who This Is For

The promise of AI productivity often comes with a significant underlying anxiety: the potential for private, sensitive data to surface where it does not belong [9][11]. Recent observations indicate that misconfigured Microsoft 365 Copilot agents and search functions can accidentally expose internal emails and files to unauthorized users within an organization [1][3][12]. While these tools are designed to streamline workflows, the rapid adoption of AI has frequently outpaced the implementation of traditional security controls [9][11].

This article provides a technical overview for Microsoft 365 administrators, IT security professionals, and end-users who are concerned about data residency and permission boundaries [2][10]. It details how specific configurations in Microsoft Copilot Studio and Copilot Chat may lead to unintended data visibility [1][3][12].

The scope of this report is limited to:

  • Security risks associated with Microsoft 365 Copilot and its integrated agents [1][9].
  • Data governance and authentication misconfigurations within the Microsoft ecosystem [1][11].
  • Methods for detecting and mitigating internal data exposure [1][3].

This article does not cover non-Microsoft AI platforms, third-party Large Language Models (LLMs), or consumer-grade AI tools outside of the Microsoft 365 environment.

TL;DR What This Means for You

A critical software defect in Microsoft 365 Copilot recently allowed the AI to bypass privacy protections and access sensitive information [1][4]. If your organization uses Copilot Chat, here is the essential information:

  • The Breach: A bug (tracked as CW1226324) allowed the AI to scan and summarize emails in Outlook's "Sent" and "Drafts" folders, even when they were explicitly marked with "confidential" sensitivity labels [1][3][9].
  • The Scope: The vulnerability bypassed established Data Loss Prevention (DLP) policies intended to block automated tools from reading protected content [3][4][9].
  • The Fix: Microsoft began rolling out a patch in early February 2026 to resolve the technical error [4][28][32]. While the immediate pathway is being closed, some organizations are still being monitored for successful remediation [7].
  • Immediate Action: Administrators should review Microsoft 365 audit logs and verify that sensitivity labels are correctly enforced across all AI-enabled workflows [28][35].
  • Long-term Outlook: Experts suggest that while this specific bug is being patched, the incident highlights a "material risk" regarding how AI assistants retrieve data across enterprise environments [3][7][10].
Concern Status Action Required
Data Exposure Confirmed for Sent/Draft folders [1][10]. Audit recent Copilot Chat summaries for sensitive data leaks [28].
DLP Bypass Resolved via software patch [4][32]. Confirm the patch status within your Microsoft 365 admin portal [3].
Compliance Potential regulatory exposure [10][34]. Review internal AI data-flow policies and risk assessments [35].

The incident serves as a reminder that generative AI integrations can create new pathways for data access that traditional security controls may not initially catch [7][34]. Continued scrutiny of workspace tools that surface internal communications is generally recommended for all IT teams [35].

Background: How Copilot Interacts with Your Data

To understand how data exposure can occur, it is essential to first understand how Microsoft 365 Copilot functions within a corporate environment. Copilot is not a standalone search engine; it is an AI assistant integrated into the applications employees use daily, such as Word, Teams, Outlook, and SharePoint [3][4].

The Role of the Large Language Model (LLM)

At its core, Copilot uses a Large Language Model (LLM). In simple terms, an LLM is a sophisticated AI trained to understand, summarize, and generate human-like text by identifying patterns in massive amounts of information [3][13].

When a user enters a prompt, Copilot does not just rely on its general training. It "indexes" or scans the organization's internal data to provide answers that are specific to that business [3][11]. This process allows the AI to summarize missing emails, create project plans based on existing files, or recap meetings from transcripts [4][6].

Understanding the Microsoft Graph

The "connective tissue" that allows Copilot to see this information is called the Microsoft Graph. This is an underlying framework that maps the relationships between users, files, emails, and calendar events [3][4].

Component Function in Copilot
Data Sources Scans Outlook, Teams, OneDrive, and SharePoint [3].
Contextual Retrieval Pulls information from meeting recaps, chats, and calendars to answer specific queries [4].
Search vs. Chat Search provides fast, targeted access to links; Chat focuses on explorational interactions [3].

Data Loss Prevention (DLP) and Permissions

Under normal operating conditions, Copilot is designed to respect existing role-based access controls (RBAC) [3]. This means a user should potentially only see information they already have permission to access through traditional folders or email threads [3].

Data Loss Prevention (DLP) refers to the set of rules and tools intended to ensure that sensitive information—such as private emails or financial records—is not lost, misused, or accessed by unauthorized users [1]. While these systems are typically robust, they rely on correct configurations. Industry observations suggest that misconfigurations, such as sharing agents with an entire organization or failing to require proper authentication, can potentially lead to unintended data exposure or "data leakage" [1][15].

How Agents Expand Access

The ecosystem has recently evolved to include Copilot agents [11]. While standard Copilot acts as an assistant, agents are purpose-built tools designed to automate specific tasks or "act" on behalf of a user [11][14]. Because these agents can be created by any employee and shared across a team, they represent a new layer of data interaction that requires careful governance to prevent unauthorized access to internal files [1][6].

Problem Explanation: The 2026 Email Exposure Bug

The incident, identified by administrators under the tracking label CW1226324, involved a critical software defect in the Microsoft 365 Copilot integration [6][11][14]. This bug allowed the AI assistant to access, read, and summarize confidential emails without explicit user permission or administrative consent [6][35]. Reports indicate the issue primarily affected Copilot Chat features within the Office productivity suite, including Outlook, Word, and Excel [6][30].

The flaw was particularly severe because it allowed the AI to bypass established Data Loss Prevention (DLP) policies [6][11][30]. These security protocols are designed to act as "brakes," preventing automated tools from processing or moving sensitive information [11][34]. Even when messages were marked with specific "Confidential" or "Highly Confidential" labels, the AI continued to scan and outline the contents for users within the same organization [30][34].

The bug specifically targeted content held within two sensitive areas of a user's mailbox:

Feature Expected Behavior Bug Behavior (CW1226324)
Confidential Labels Block AI scanning and processing [30] Labels were ignored by the retrieval system [11][30]
DLP Policies Prevent unauthorized data retrieval [34] Policies were bypassed during the chat workflow [6][25]
Drafts/Sent Folders Restricted from automated summarization [28] Contents were summarized without user consent [11][28]

This exposure has raised substantial concerns regarding corporate confidentiality and the long-term trust in enterprise generative AI tools [21][35]. Because the affected folders often contain business contracts, legal correspondence, and personal medical information, unauthorized processing creates immediate compliance and regulatory risks [11][34][35]. While current evidence suggests there was no "cross-tenant" leakage—meaning data was not shared between different companies—the internal exposure of restricted data remains a significant security failure [30][34].

The behavior was first reported by customers on January 21, 2026 [9][14]. It reportedly persisted for several weeks before Microsoft acknowledged the issue and began rolling out a fix in early February [6][13][35]. Industry analysts suggest that this event underscores why many large organizations now cite AI as a material risk in regulatory filings, as the integration of these tools creates new pathways for data to circumvent traditional security controls [9][10][21].

Why This Happens: Root Causes and Analysis

The exposure of sensitive email data through Microsoft 365 Copilot and Copilot Studio is not the result of a single failure, but rather a combination of technical bugs and configuration oversights [6][13]. While generative AI offers high productivity, its ability to interact with organizational data creates new pathways that traditional security controls may not always monitor [6].

Confirmed Root Causes

According to technical reports and official security bulletins, the following factors contributed to the data exposure:

  • Code-Level Retrieval Errors: A specific software defect in the Office and Copilot integration allowed the assistant to access folders it was intended to ignore [14][15]. Specifically, a code issue permitted the AI to pick up items from a user’s Sent Items and Drafts folders, even when those items were protected by security protocols [10][14].
  • DLP and Sensitivity Label Bypasses: Data Loss Prevention (DLP) policies and Sensitivity Labels are designed to act as "brakes" to prevent unauthorized data movement [13]. However, a bug in the Copilot Chat workflow caused the system to incorrectly process "Confidential" emails, summarizing them despite active DLP rules that should have blocked such actions [9][11].
  • Author Authentication and Privilege Escalation: In Copilot Studio, agents may be configured to use "Author Authentication" (maker credentials) [2][5]. In this state, the agent operates with the permissions of the person who created it rather than the end-user [2]. If the creator has high-level access, any user interacting with that agent could potentially access sensitive data through the creator’s identity [2][6].
  • Broad Sharing and Lack of Authentication: Many agents are shared with entire organizations or broad security groups to save time during deployment [1][3]. When these agents are also configured to require no authentication, they can behave like public entry points into internal organizational logic and data stores [1].

Analysis of Technical Vulnerabilities

The failure point often occurs during the retrieval phase rather than the generation phase [13]. While the AI is capable of following instructions, the "retrieval-time" checks—which verify if a user has permission to see a specific document or email—can be misapplied [13].

Vulnerability Factor Impact on Security
Generative Orchestration May allow an orchestrator to determine recipients at runtime, potentially leading to data exfiltration if prompt instructions are weak [3][4].
Dormant Components Unused agents or "draft" versions often lack active ownership and may use outdated security standards, creating "blind spots" for attackers [2][5].
HTTP Request Actions Direct requests to non-standard ports or insecure schemes can bypass the governance and identity controls provided by standard connectors [1][4].

Potential Ecosystem Challenges

Industry analysts and technical documentation suggest that the complexity of the Microsoft 365 ecosystem contributes to these risks. Real-time permission syncing across the Microsoft Graph, mailboxes, and documents is a massive undertaking [13].

It appears that when new AI features are integrated into legacy environments, the sheer speed at which the assistant touches different data stores can surface information that users were never intended to see [13][15]. Experts suggest that organizations should treat these AI integrations as fundamental changes to their data flows rather than simple software updates [12][14].

Evidence and Reality Check

The risks associated with AI-driven data exposure are not merely theoretical. Recent disclosures from major technology providers and independent security researchers confirm that these vulnerabilities are currently being observed and mitigated in enterprise environments [1][12].

Confirmed Technical Incidents

Microsoft has officially acknowledged a specific configuration bug, tracked by administrators as CW1226324, which affected how Microsoft 365 Copilot handled sensitive communications [12]. According to official reports, the AI incorrectly processed email messages that were explicitly marked with a confidential label [12].

This incident revealed several critical weaknesses in automated data protection:

  • Bypass of DLP Policies: The bug allowed the AI to summarize content even when customers had Data Loss Prevention (DLP) policies configured to prevent such ingestion [12].
  • Scope of Exposure: The vulnerability affected both draft and sent email messages starting in early January 2026 [12].
  • Correction Timeline: Microsoft began rolling out a fix for this specific issue in February 2026, though the total number of affected enterprise customers remains unconfirmed [11][12].

Research into AI-Based Data Exfiltration

Beyond internal software bugs, external security analysts have demonstrated that AI assistants can be intentionally manipulated for data theft. Research from Check Point highlights a technique codenamed AI as a C2 proxy, which was successfully demonstrated against Microsoft Copilot and xAI Grok [6][13].

"AI assistants are no longer just productivity tools; they are becoming part of the infrastructure that malware can abuse," warned security experts regarding the rise of stealthy, AI-assisted malware operations [8][13].

Analysts report that threat actors can hide malicious traffic "in plain sight" by encoding stolen data into URLs and asking the AI to summarize them [5][13]. Because the traffic appears to be a legitimate interaction with a trusted AI service, it often evades traditional security monitoring tools [5][15].

Observed Operational Risks

In addition to specific exploits, Microsoft’s security research team has identified several common "misconfigurations in the wild" that lead to data leakage [1][3]. These are not hypothetical bugs but actual settings found in active business environments:

Risk Category Security Impact Observed Reality
Broad Sharing Unintended access Agents frequently shared with entire organizations without access boundaries [1][4].
No Authentication Public exposure Agents found running in default states that allow anonymous access via a simple link [4].
Author Authentication Privilege escalation Agents acting on behalf of the creator's high-level permissions rather than the end user [14].

These findings suggest that the integration of AI into corporate workflows often outpaces the implementation of necessary security governance [3][14]. Experts predict that as AI agents become more integrated into operational systems, these exposure points will likely become more frequent targets for advanced threat actors [3][6].

How to Check if You Are Affected

Identifying whether your organizational data has been exposed through Microsoft Copilot Studio or Microsoft 365 Copilot involves auditing both administrative configurations and individual user activity. Potential exposure often stems from misconfigured AI agents or broad sharing settings that bypass intended data boundaries [1][8].

1. Audit Agent Sharing and Authentication

The most common point of exposure occurs when agents are shared too broadly or are configured to operate without authentication [1][4]. Administrators should review the status of all published agents to ensure they are not accessible to the entire organization or external entities unintentionally.

  • Check Sharing Settings: Verify if any agents are shared with "Everyone" or broad security groups, as this expands the attack surface for unauthorized data access [1][8].
  • Verify Authentication Requirements: Identify agents that have authentication deactivated or set to "only prompt on demand," which can turn an internal agent into a public entry point for organizational data [1].

2. Utilize Advanced Hunting Queries

For organizations using Microsoft Defender, security teams can use Advanced Hunting to detect specific misconfigurations that might lead to data leakage [1][2].

  • Run Community Queries: Navigate to the Security portal > Advanced hunting > Queries > AI Agent folder to run specific detections [1].
  • Identify Risky Actions: Look for high-risk indicators such as "AI Agents – No Authentication Required" or "AI Agents – Sending email to external mailboxes" to pinpoint potential exfiltration paths [1].

3. Review Copilot Interaction History

Individual users can monitor how the AI is interacting with their data by reviewing their interaction history and the summaries generated by the tool [4].

  • Review Summaries: Check recent Copilot Chat summaries of meetings, emails, and chats. If Copilot summarizes information that a user should not have access to, it may indicate a permissions or sensitivity label failure [4].
  • Monitor Search Activity: Use the Copilot search function to test if sensitive files are retrievable via simple prompts. For example, a prompt such as "Find the file shared with me within the last six months" can reveal if files are appearing in search results unexpectedly [3].

4. Consult the Data Compliance Manager Role

Microsoft provides specific tools for data compliance managers to audit how Copilot respects existing role-based access controls (RBAC) [3].

  • Explain Data Access: Authorized personnel can use sample prompts within Copilot to generate reports on what data is accessible. A prompt such as "Explain what data Microsoft 365 Copilot can access within my organization" can help clarify how residency and retention are being handled [3].
  • Audit Permissions: Ensure that Copilot is respecting existing sensitivity labels. If a user can prompt Copilot to summarize a folder protected by restrictive sensitivity labels, the governance settings may require adjustment [3].

5. Check for Dormant or Orphaned Agents

Risks often persist in "blind spots," such as agents created for testing that were never deactivated [13].

  • Identify Dormant Assets: Look for agents that have not been modified or invoked in over 30 days. These "dormant" agents may use outdated logic or insecure connections that do not meet current security standards [1][13].
  • Review Author Authentication: Check if agents are running using "Author" (Maker) authentication. In this state, any user interacting with the agent inherits the creator's potentially elevated permissions, leading to privilege escalation [1][13].

What You Can Do: Solutions and Mitigation

Organizations can significantly reduce the risk of unintended data exposure by addressing specific configuration gaps in their Copilot Studio agents and AI workflows [1][2]. Proper security posture requires a combination of immediate technical restrictions and long-term governance changes [1].

Short-Term Options: Immediate Security Hardening

For organizations needing to secure their environment quickly, the following actions address the most common misconfigurations observed in the wild [1]:

  • Restrict Sharing Scopes: Organizations should move away from sharing agents with the entire organization or broad security groups [1][3]. Access should be limited only to the specific users or roles that require the agent's functionality to perform their duties [3][6].
  • Enforce Mandatory Authentication: Agents should be configured to require authentication by default [1]. Allowing unauthenticated access or "authentication on demand" can potentially turn an internal agent into a public entry point for organizational data [6].
  • Audit HTTP and Email Actions: Security teams can review agents that use HTTP Request actions or email-based outputs [1]. Restricting these to secured connectors rather than direct URLs or externally controlled email fields helps prevent data exfiltration and insecure communications [9].
  • Deactivate Dormant Agents: Identifying and removing agents or connections that have not been modified or invoked in over 30 days reduces the hidden attack surface [1][7]. These orphaned or unused assets often lack active ownership and may contain outdated security logic [7].

Long-Term Options: Strategic Governance

Establishing a sustainable AI security model involves moving toward systemic controls that minimize human error during the creation process [1].

  • Transition to a "Least Privilege" Model: All agents should ideally operate using the end-user's identity rather than the creator's (author) authentication [1][7]. When agents use author authentication, every user inherits the maker's permissions, which potentially allows for privilege escalation if the maker has high-level access [7].
  • Govern Model Context Protocol (MCP) Tools: Organizations should implement strict oversight for MCP tools that connect agents to internal data tables [3][7]. These tools can create undocumented access paths if they are configured with maker credentials instead of user-specific permissions [7].
  • Implement Generative Orchestration Safeguards: For agents using generative orchestration, it is critical to provide clear, restrictive instructions [1][3]. Lacking specific instructions can lead to "behavior drift" where the agent performs unintended actions based on user prompts [3].

Comparison of Mitigation Strategies

Strategy Focus Area Security Impact
Access Control Restricting sharing scopes Reduces the available attack surface [3][6].
Identity Management Disabling Author Authentication Prevents privilege escalation and maintains separation of duties [7].
Resource Governance Monitoring dormant/orphaned agents Eliminates unmanaged entry points into the network [7].
Data Protection Restricting HTTP and Email actions Minimizes the risk of data exfiltration via prompt injection [9].

Risks & Limitations

While these steps significantly improve security, no configuration is entirely risk-free [1]. Security teams must balance restrictive controls with user productivity to prevent "shadow AI" usage [14]. Furthermore, because the Copilot ecosystem is evolving rapidly, configurations that are secure today may require updates as new capabilities, such as advanced Agent Mode features, are released [11][14].

It is generally recommended that only the creator maintains edit control during the initial phases of agent deployment to prevent unauthorized logic changes [8][15]. If a configuration becomes too complex for internal teams to manage, it is typically more cost-effective to consult an expert than to remediate a large-scale data leak later.

Risks, Limits, and When to Stop

Software patches and updates are essential for maintaining system integrity, but they may not eliminate 100% of AI-related risks [1][6]. While Microsoft investigates the extent of recent vulnerabilities, it is generally observed that the scope of impact for such bugs can change as more information becomes available [10].

The Persistence of Configuration Risks

Even when technical bugs are resolved, organizational misconfigurations can leave environments vulnerable. Research indicates that several common issues often persist after initial setup:

  • Broad Sharing: Agents shared with an entire organization or large groups significantly expand the attack surface [1][2].
  • Authentication Gaps: Agents that do not require authentication or were left in a default "test" state act as public entry points into internal data [1][3].
  • Dormant Assets: Unpublished drafts, unused actions, or orphaned agents without active owners often fall outside normal security monitoring [1][4].

External Threats and Prompt Abuse

Advanced threats like cross-prompt injection (XPIA) can potentially bypass standard safeguards to exfiltrate data via email or other channels [1][5]. Experts warn that AI assistants can be hijacked to act as stealthy transport layers for malicious traffic, hiding command-and-control (C2) operations within legitimate AI requests [8][15].

Risk Category Potential Security Impact
Generative Orchestration Behavior drift or unintended actions due to lacking instructions [1][2].
Author Authentication Privilege escalation when an agent uses the creator's high-level permissions [1][4].
HTTP Request Actions Governance bypass through insecure communications or non-standard ports [1][3].

When to Stop and Seek Professional Help

Attempting a manual cleanup of a widespread data exposure can be complex and may inadvertently overlook hidden attack paths. It is recommended to stop using the service and consult IT security professionals if any of the following occur:

  1. Suspected Data Leakage: If confidential communications appear to have been repurposed or surfaced to unauthorized users [7][10].
  2. Unidentified Ownership: If "orphaned" agents with disabled owners are discovered running in the environment [1][2].
  3. Active Exploitation: If there are signs of AI-driven implants or automated triage being performed by unauthorized entities [13][15].

IT teams are advised to treat new AI integrations as fundamental changes to data flows [7]. This includes updating risk assessments and ensuring that incident response plans specifically cover AI-related failures [7]. If you are unsure about the security posture of an AI agent, it is typically safer to disable the tool until a full audit can be performed.

FAQ

Was the Microsoft 365 Copilot email bug fixed?

Microsoft confirmed that a fix for the vulnerability, tracked by administrators as CW1226324, began rolling out in early February 2026 [15]. The bug previously allowed Copilot Chat to summarize draft and sent emails labeled as "confidential" even when data loss prevention policies were active [15].

Does Copilot learn from my private data?

Microsoft 365 Copilot is designed to respect existing role-based access controls and permissions within an organization [10]. However, this specific bug incorrectly processed sensitive information for several weeks, bypassing intended restrictions [15]. Under normal operating conditions, the system is intended to use work data as a reference without compromising established security labels [10][15].

Can administrators control how Copilot accesses data?

Yes, administrators have access to an admin portal where they can manage features and customize how information is retrieved [10]. Additionally, security tools like Microsoft Defender can be used to detect misconfigured agents that might lead to unintended access or data leakage [13]. Organizations can also implement data loss prevention policies, though the recent bug highlights that these may not be 100% effective during technical failures [15].

How do I know if my confidential emails were exposed?

Administrators can monitor the issue through the service health dashboard using the identifier CW1226324 [15]. The exposure specifically affected paying Microsoft 365 customers using the AI-powered chat features within Office software products like Outlook, Word, and Excel [1][15]. It appears limited to messages where a "confidential" label had been manually or automatically applied [15].

Can I build my own agents to handle data more securely?

Employees can create lightweight agents to automate repetitive tasks and keep team data aligned [4][5]. However, security researchers note that agents must be configured correctly to require authentication and avoid "dormant connections" that could hide an attack surface [13]. Currently, only the creator of an agent maintains edit control over its functions [4][9].

Summary and Key Takeaways

The integration of AI agents like Microsoft Copilot Studio into organizational workflows introduces powerful automation but also creates significant security blind spots [4]. Recent observations indicate that small, often well-intentioned configuration choices—such as sharing agents too broadly or disabling authentication for testing—can inadvertently turn into critical entry points for data leakage [3][5].

Key Insights

  • Authentication Gaps: Agents configured without mandatory authentication or those using "author (maker) authentication" can lead to unauthorized data access and privilege escalation [1][5].
  • Dormant Risks: Unused or "orphaned" agents and unpublished drafts often lack active ownership and may contain outdated logic or sensitive connections that do not meet current security standards [1][3].
  • Orchestration Vulnerabilities: Agents using generative orchestration without strict instructions are susceptible to prompt injection attacks, potentially allowing threat actors to exfiltrate internal data via email [2][6].
  • Governance Requirements: Organizations are encouraged to treat new AI integrations as fundamental changes to data flows, necessitating updated risk assessments and tightened controls over workspace tools [15].

Core Recommendations

Action Item Goal
Audit Permissions Ensure agents are not shared with the entire organization unless necessary [3].
Enforce Identity Use end-user authentication rather than maker credentials to maintain the principle of least privilege [1].
Monitor Activity Regularly identify and remove dormant agents or unused connectors to reduce the attack surface [2].
Secure Communications Restrict HTTP request actions and email-based data outputs to prevent unauthorized exfiltration [5][6].

As AI agents become a core part of operational systems, the exposure they create becomes both easier for threat actors to find and more dangerous for organizations to ignore [4]. Managing these risks effectively requires a shift from viewing AI as a simple plug-and-play tool to treating it as a complex infrastructure component that requires constant governance [15].

If you are unsure about the security posture of your AI environment, it is usually cheaper to ask for a professional security audit than to fix a significant data breach later.

Quellen

[1] Top 10 actions to build agents securely with Microsoft Copilot Studio | Micro...

[2] What’s new in Power Platform: February 2026 feature update - Microsoft Power ...

[3] Picking the right Copilot for the job: Tips from our experience at Microsoft ...

[4] A day in the life of a Microsoft employee using Copilot - Inside Track Blog

[5] The Hacker News - Google News

[6] Microsoft says Office bug exposed customers' confidential emails to Copi...

[7] Researchers Show Copilot and Grok Can Be Abused as Malware C2 Proxies

[8] Experts warn Copilot and Grok can be hijacked to spread malware

[9] Copilot Chat bug bypasses DLP on

[10] Weekly Recap: Outlook Add-Ins Hijack, 0-Day Patches, Wormable Botnet & AI...

[11] Copilot bug allows 'AI' to read confidential Outlook emails

[12] New Chrome Zero-Day (CVE-2026-2441) Under Active Attack — Patch Released

[13] Critical Flaws Found in Four VS Code Extensions with Over 125 Million Installs

[14] Capita taps Microsoft Copilot to untangle UK pensions mess

[15] European Parliament blocks AI on lawmakers' devices, citing security ris...

[16] Simform Scales Microsoft Practice with $3mn investment; Targets IP, and Co-Se...

[17] /C O R R E C T I O N -- Educate 360/

[18] Sifted Launches SiftedAI Copilot, an Agentic AI Platform Overlay Delivering A...

[19] Cognizant to provide strategic technology services for Wallenius Wilhelmsen

[20] AI Is Rewriting Healthcare Publishing -- Admanager, Powered by Doceree, Launc...

[21] Microsoft Copilot Bug Exposed Customer Emails to AI

[22] The Copilot Problem: Why Internal AI Assistants Are Becoming Accidental Data ...

[23] RoguePilot: Exploiting GitHub Copilot for a Repository Takeover

[24] Microsoft Confirms Copilot AI Security Flaw Exposing Confidential Emails | Uk...

[25] Microsoft Copilot read confidential emails without permission

[26] Uncovering the Sophisticated Phishing Campaign Bypassing M365 MFA

[27] What Are Copilot Agents? Intro to Copilot Studio Lite vs. Full​- C5 Insight

[28] Why did Microsoft 365 Copilot summarize confidential emails?

[29] I finally removed every trace of Copilot and Recall from Windows — here'...

[30] Microsoft Confirms Office Bug Exposed Emails To Copilot

[31] Microsoft Purview’s Adaptive DLP Brings Scalable Control to Copilot

[32] Microsoft says Copilot was summarizing confidential emails without permission

[33] Copilot & Grok AI Vulnerable to Prompt Attacks, Researchers Claim

[34] Microsoft Copilot Read Confidential Emails Without Consent

[35] How did Microsoft's Copilot expose emails?

[36] Microsoft 365 Copilot Flaw Allows AI Assistant to Summarize Sensitive Emails

[37] Microsoft Teams Introduces AI Workflows Powered by Microsoft 365 Copilot for ...

[38] FTC digs deeper into Microsoft’s bundling and licensing practices

[39] How to Help Copilot Encourage Good Database Standards - Brent Ozar Unlimited®

[40] Microsoft Teams With AI Workflows Use Microsoft 365 Copilot to Automate Tasks...

[41] Microsoft 365 Pricing Changes in 2026: Plan Increases and Exceptions

[42] A new approach for GenAI risk protection

[43] How to do AI analysis you can actually trust

[44] EU Parliament blocks AI tools over cyber, privacy fears

[45] Cision - Global Cloud-Based Communications and PR Solutions Leader

[46] PR Newswire for Agency Partners

[47] PR Newswire | LinkedIn

[48] Cision - Global Cloud-Based Communications and PR Solutions Leader

[49] XFN 1.1 profile

[50] The Hacker News

[51] fonts.googleapis.com

[52] The Hacker News | LinkedIn

[53] The Hacker News

Relevant Services

More from the Blog

Brauchen Sie Hilfe?

Wir reparieren Ihren PC oder Laptop schnell und zuverlässig.

Jetzt Reparatur anfragen