How to use this reference
Define output-safety invariants for every observable intermediate state, not just the completed answer. Keep generated content inert until applicable policy checks have passed. Review rendering and data-disclosure boundaries together; a clean final display is insufficient evidence that no earlier side effect occurred.
Before reading
- Browser rendering, sanitization and content-security-policy concepts
- AI output trust boundaries and ordering of security checks
Context and limits
- The reported conditions require affected enterprise search, relevant content accessible to the victim and the victim opening a supplied link. This resource isolates the timing boundary; disclosure also depended on additional weaknesses.
- Wider indexed-content exposure depends on the victim's access. Account takeover is a proposed consequence, not a demonstrated result in this record.
- Varonis says Microsoft patched the issue. The exact fix date and implementation are not established by the reviewed evidence.
- The Microsoft CNA record corroborates network information disclosure requiring user interaction and identifies an exclusively hosted service. It supplies no usable affected-version range. Its CWE-77 classification does not establish operating-system command execution.
- The CNA's June 4, 2026 public-disclosure date and September 24 update date describe the vulnerability record, not this article's publication or remediation chronology.
- No individual award amount or affected-victim count is established. This educational record grants no testing authorization.
Sources and provenance
- SearchLeak: How We Turned M365 Copilot Into a One-Click Data Exfiltration Weapon Varonis Threat Labs · reviewed 2026-10-04
- Varonis Blog catalog: SearchLeak entry Varonis · reviewed 2026-10-04
- Microsoft CNA record for CVE-2026-42824 Microsoft via CVE Program · reviewed 2026-10-04
Record reviewed 2026-10-04. Snapshot d5550c789111. Open the complete JSON contract.