This is a research prototype. The data and analyses are preliminary and not yet validated — we'd welcome your .
A security researcher discovered a vulnerability in Microsoft 365 Copilot that allowed users to access and summarize files without triggering audit log entries. This flaw could be exploited by malicious insiders to exfiltrate data undetected, posing significant risks to organizations requiring strict compliance and security logging. Microsoft fixed the issue in August 2025 but declined to issue a CVE or notify customers, citing the vulnerability's 'important' classification.
A vulnerability in Microsoft 365 Copilot reportedly allowed users to access and summarize files without generating audit log entries, allegedly undermining traceability and compliance. Security researcher Zack Korman disclosed the issue to Microsoft, which reportedly classified it as "important" and fixed it on August 17, 2025, but reportedly chose not to notify customers or assign a CVE.
Domain classification, causal taxonomy, severity scores, and national security assessments were LLM-classified and may contain errors.
Vulnerabilities that can be exploited in AI systems, software development toolchains, and hardware, resulting in unauthorized access, data and privacy breaches, or system manipulation causing unsafe outputs or behavior.
The incident involves a security vulnerability in Microsoft 365 Copilot that allowed users to bypass audit logging, creating a significant security flaw that could be exploited for unauthorized data access.
AI system
Due to a decision or action made by an AI system
Unintentional
Due to an unexpected outcome from pursuing a goal
Post-deployment
Occurring after the AI model has been trained and deployed
The risk stems from an unintentional security vulnerability in the deployed Microsoft 365 Copilot system, where specific user prompts caused the AI to bypass standard audit logging.
A security researcher discovered a vulnerability in Microsoft 365 Copilot that allowed users to bypass audit logging when accessing files. While representing a minor information security risk due to potential undetected insider data exfiltration, the issue has been patched by Microsoft, and there are no reports of active exploitation by hostile actors.
Microsoft 365 Copilot is described as an AI assistant that helps users perform daily tasks, interact with services, and summarize files, operating under direct human instruction.
Risk Level 3: Limited Risk. The system is a general-purpose AI assistant/chatbot, which is subject to transparency obligations under the EU AI Act.
Microsoft 365 Copilot is primarily designed as an interactive chatbot and writing assistant that helps users summarize documents and interact with office services.
The report provides a clear, first-hand technical account of the vulnerability, including reproduction steps, expected vs. actual behavior, and communication logs with Microsoft.
None
Technology
Prompt Instructions Bypass Logs
Asking Copilot to omit links resulted in no audit log entry being recorded.
Truncation of Chat History
Copilot truncates chats, making historical verification of bypasses difficult.
Data Inputs
Unsanitized Prompt Inputs
User prompts could directly manipulate the generation of audit logs.
Human Factors
Accidental Bypass Triggering
Users could trigger the audit log bypass unintentionally during normal work.
Malicious Insider Exploitation
Insiders could intentionally use Copilot to steal files without detection.
Process and Methods
Inconsistent Vulnerability Process
Microsoft deviated from its own security reporting and status tracking guides.
Flawed CVE Assignment Rules
Policies prevented CVE assignment because the fix was pushed automatically.
Management
Failure to Address Prior Report
Zenity CTO reported the issue a year prior, but management did not fix it.
Non-Disclosure Decision
Management chose not to notify customers or publicize the vulnerability.
Downplaying Vulnerability Severity
Classifying the issue as 'important' rather than 'critical' to avoid CVE.
Regulatory Environment
Compliance Verification Gap
Regulated entities like HIPAA users relied on incorrect logs for compliance.
AI system involved
Identified by automated extraction; can differ from the alleged developers above.
