๐ŸŒRecentcollected in 60m

SearchLeak vulnerability exposes Microsoft 365 Copilot data

SearchLeak vulnerability exposes Microsoft 365 Copilot data
PostLinkedIn
๐ŸŒRead original on The Next Web (TNW)

๐Ÿ’กCritical security flaw in Microsoft 365 Copilot allows data theft via a single link. Essential for enterprise security.

โšก 30-Second TL;DR

What Changed

Vulnerability chain identified in Microsoft 365 Copilot Enterprise Search

Why It Matters

This vulnerability highlights the critical need for robust data access controls within enterprise AI agents. Organizations must audit their Copilot search permissions to prevent unauthorized data exposure.

What To Do Next

Audit your Microsoft 365 Copilot search permissions and restrict access to sensitive data sources immediately.

Who should care:Enterprise & Security Teams

๐Ÿง  Deep Insight

Web-grounded analysis with 12 cited sources.

๐Ÿ”‘ Enhanced Key Takeaways

  • โ€ขThe 'SearchLeak' vulnerability, assigned CVE-2026-42824 with a critical severity rating by Microsoft, is a three-stage attack chain combining a Parameter-to-Prompt (P2P) Injection, an HTML Rendering Race Condition, and a Content Security Policy (CSP) Bypass via Bing Server-Side Request Forgery (SSRF).
  • โ€ขThe exploit allowed attackers to silently exfiltrate sensitive data such as MFA codes, email messages, calendar details, and private organizational files from a victim's mailbox, calendar, SharePoint, and OneDrive accounts.
  • โ€ขMicrosoft has already remediated the 'SearchLeak' vulnerability on its backend, meaning no user action is required to mitigate this specific threat.
  • โ€ขThe vulnerability leveraged a relatively new class of AI-specific weakness (Parameter-to-Prompt Injection) in conjunction with classic web security bugs, highlighting how AI systems can create new pathways to exploit older vulnerabilities in impactful ways.
  • โ€ขTraditional anti-phishing filters and URL filtering tools were unlikely to detect the malicious link because it pointed to a legitimate microsoft.com domain, making the attack highly stealthy.
๐Ÿ“Š Competitor Analysisโ–ธ Show

Microsoft 365 Copilot Competitor Analysis

Feature/AspectMicrosoft 365 CopilotOpenAI ChatGPT Enterprise / TeamsAnthropic Claude for EnterpriseGoogle Gemini for WorkspaceGoSearchGleanCoworker AI
Primary FocusAI assistant embedded across Microsoft 365 suiteB2B conversational AI with enhanced securityLeading competitor to OpenAI, strong Amazon backingMultimodal AI with native Google Workspace integrationAgentic Enterprise AI Search across entire tech stackEnterprise search platform across internal systemsAI that both searches and executes actions across mixed stacks
Data IntegrationMicrosoft Graph (M365 data: Word, Excel, Outlook, Teams, SharePoint, OneDrive)Enterprise data (specifics depend on integration)Enterprise data (specifics depend on integration)Google Workspace data100+ integrations (Slack, Jira, Confluence, Salesforce, etc.)100+ integrations (Slack, Box, Google Drive, OneDrive, SharePoint, etc.)40+ native integrations (Salesforce, Slack, Jira, Google Drive, etc.)
Security PostureEnterprise-grade data privacy and security, honors existing M365 controlsEnhanced security, collaboration, user managementStrong security, guardrailsEnterprise-grade data privacy and securityHybrid federated and indexed architecture, permission-aware answersRobust compliance and end-to-end data security, guardrailsEnterprise-grade security
Key CapabilitiesDocument drafting, data analysis, email summarization, team communication, unified searchConversational AI, team workflows, browser tasksAutonomous multi-step tasks on local files/appsDeep Research, real-time web accessUnified AI search and action, knowledge centralizationPowerful search, generative AI features, cites sourcesCross-tool search and execution (e.g., update CRM, create Jira tickets)
Pricing (per user/month)Bundles start at $22 (promotional), add-on for existing M365 subscribersNot specified, B2B versionNot specified, B2B versionGemini Advanced: $20Custom enterprise pricing (typically $10-15)Custom enterprise pricing (typically $10-15)$30 (no M365 license required)
LimitationsPrimarily limited to Microsoft 365 ecosystem; challenges with mixed tech stacksSearch-only for some aspects (e.g., Glean)Search-only, not for market/investment research

๐Ÿ› ๏ธ Technical Deep Dive

  • Parameter-to-Prompt (P2P) Injection: The vulnerability chain began by exploiting how Microsoft 365 Copilot Enterprise Search processes the 'q' URL parameter. Instead of treating the content of this parameter solely as a search query, Copilot interpreted it as executable instructions, allowing an attacker to inject malicious prompts.
  • HTML Rendering Race Condition: The second stage involved a timing vulnerability. When Copilot streamed its output, an injected <img> tag within the AI's response would fire its request to an external server before the output sanitizer could wrap the raw HTML in <code> blocks, which would neutralize it. This allowed the image request, containing exfiltrated data, to be sent before being blocked.
  • CSP Bypass via Bing SSRF: The final stage exploited Bing's image-search endpoint, which was allowlisted in the Content Security Policy (CSP). This endpoint performed a server-side fetch to an attacker-controlled URL. The <img> tag, triggered by the race condition, directed Bing to fetch the attacker's URL, which included the stolen data, effectively bypassing the page's CSP and exfiltrating the information.

๐Ÿ”ฎ Future ImplicationsAI analysis grounded in cited sources

AI-specific vulnerabilities will increasingly combine with traditional web flaws.
The 'SearchLeak' vulnerability demonstrates a growing trend where novel AI-specific weaknesses, like prompt injection, are chained with older, well-understood web security bugs (e.g., race conditions, SSRF) to create potent new attack vectors.
Traditional security tools will struggle to detect sophisticated AI-driven exploits.
Because the malicious link in 'SearchLeak' originated from a legitimate microsoft.com domain, conventional anti-phishing and URL filtering solutions were ineffective, necessitating new detection methods for AI-powered platforms.
Continuous security auditing and 'red-teaming' of AI systems will become critical.
The discovery of 'SearchLeak' and previous vulnerabilities like 'Reprompt' and 'EchoLeak' underscore the need for ongoing, specialized security research to identify and mitigate complex, chained vulnerabilities in AI assistants and enterprise search tools.

โณ Timeline

2025-03
Microsoft 365 Copilot generally available for enterprise customers
2025-05
Aim Security discloses 'EchoLeak' (CVE-2025-32711), a zero-click Copilot data-leak bug
2025-12
Varonis Threat Labs discovers 'Reprompt' vulnerability in Copilot Personal
2026-05
Microsoft applies server-side fixes for the 'SearchLeak' vulnerability (CVE-2026-42824)
2026-06
Varonis Threat Labs publicly discloses 'SearchLeak' vulnerability in Microsoft 365 Copilot Enterprise Search

๐Ÿ“Ž Sources (12)

Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.

  1. varonis.com
  2. bleepingcomputer.com
  3. thehackernews.com
  4. thenextweb.com
  5. vellum.ai
  6. coworker.ai
  7. gosearch.ai
  8. alpha-sense.com
  9. microsoft.com
  10. microsoft.com
  11. microsoft.com
  12. letsdatascience.com
๐Ÿ“ฐ

Weekly AI Recap

Read this week's curated digest of top AI events โ†’

๐Ÿ‘‰Related Updates

AI-curated news aggregator. All content rights belong to original publishers.
Original source: The Next Web (TNW) โ†—