Executive Summary
Reported attacks and vulnerabilities show that untrusted webpage content or crafted links can cause AI systems to disclose user information, conversation prompts, connected-application data, or session data. Teams operating AI agents and LLM applications with browsing or application integrations should prioritize prompt-injection testing and access controls.
What Happened
The coverage reports an attack technique that could cause Grok to transmit user information and conversation prompts to an attacker-controlled server after summarizing a webpage. It also covers Microsoft Copilot vulnerabilities that could allow a crafted link to retrieve data from connected applications and the victim’s session.
Affected Products
- Grok
- Microsoft Copilot
Vulnerability / Attack Technique
Indirect prompt injection through webpage content or crafted links
Technical Mechanism
After summarizing a webpage, Grok could be induced to transmit user information and conversation prompts to an attacker-controlled server. In Microsoft Copilot, a crafted link could retrieve data from connected applications and the victim’s session.
AI Security Category
LLM01:2025 Prompt Injection
Part of the Prompt Injection topic.
Relevant Frameworks
RealGround Analysis
These reports matter because AI agents that browse content or access connected applications may treat attacker-controlled inputs as instructions and operate with the user’s privileges. RealGround recommends validating tool and link handling, limiting agent access to connected data, and testing whether prompts or session information can leave approved boundaries.
Recommended Controls
- Treat webpage content, links, and retrieved documents as untrusted data rather than executable instructions.
- Apply least-privilege permissions and per-request authorization for connected applications and session data.
- Require user confirmation before external transmission, sensitive-data access, or high-impact tool actions.
- Use egress controls and monitoring to block or alert on transmissions to unapproved destinations.
- Conduct AI red-team tests for indirect prompt injection, data retrieval, prompt leakage, and exfiltration paths.
Related RealGround Service
AI Agent Business Logic Audit · or start with the free AI Agent Security Scanner.
Sources
The Hacker News: https://thehackernews.com/search/label/Prompt%20Injection?max-results=20&m=1&hl=ja_JP
Summary and analysis written by RealGround from the source above; last updated 2026-10-20.