Researchers Detail ChatGPhish, a Phishing Risk in ChatGPT Web Summaries

Summary
Researchers say ChatGPhish can make ChatGPT render attacker-controlled links and auto-fetch images from summarized pages, potentially exposing user metadata and presenting phishing content. The article also surveys attacks targeting AI coding agents and other AI tools.
Key points
- Permiso Security says ChatGPhish abuses ChatGPT’s handling of Markdown links and images from third-party pages it summarizes.
- Auto-fetched attacker-hosted images could expose a user’s IP address, User-Agent, and Referer; attacker-controlled links, fake alerts, or QR codes could also appear in the response.
- SymJack can use a malicious repository and symlinked configuration file to cause an AI coding assistant to run attacker code after restart.
- TrustFall describes a malicious repository that can auto-approve and launch an MCP server with the developer’s privileges after a folder-trust prompt.
- The article also describes attacks affecting Claude, BrowserOS, and other AI systems; it says WebPromptTrap was patched in BrowserOS 0.32.0.
Article Details
- Vulnerability Types
- Third-party Markdown links and image URLs rendered as trusted, clickable content in ChatGPT summaries
- Indirect prompt injection through web pages summarized by ChatGPT
- Exploitation Status
- unknown
- Exploit Availability
- unknown
- Patch Status
- unknown
CVE
CVE-2026-25592microsoft · semantic-kernelA set of vulnerabilities in Microsoft Semantic Kernel (CVE-2026-25592 and CVE-2026-26030) that could turn a prompt injection into host-level remote code execution.CVE-2026-26030microsoft · semantic-kernelA set of vulnerabilities in Microsoft Semantic Kernel (CVE-2026-25592 and CVE-2026-26030) that could turn a prompt injection into host-level remote code execution.
People
Vendors
Products
ChatGPTResearchers found that ChatGPT can be abused through a vulnerability called ChatGPhish, where Markdown links and images from summarized web pages are rendered as trusted, clickable content that can leak data and enableMicrosoft Semantic KernelA set of vulnerabilities in Microsoft Semantic Kernel (CVE-2026-25592 and CVE-2026-26030) that could turn a prompt injection into host-level remote code execution.
