Return to Threats

AI Security Incident Database – DuneSlide and Other Agent Exploits

Wraith.sh 2026-08-21 AI agent abuse Critical

Executive Summary

DuneSlide is a pair of critical Cursor IDE flaws (CVE-2026-50548, CVE-2026-50549, both CVSS 9.8) that turn zero-click prompt injection into full OS remote code execution through how the agent's tools handle untrusted content.

What Happened

Wraith.sh’s AI security incident database describes DuneSlide, two critical vulnerabilities in the Cursor IDE (CVE-2026-50548 and CVE-2026-50549, both CVSS 9.8) that turn zero‑click prompt injection into full operating‑system remote code execution by abusing how the agent’s tools handle untrusted content.[13] The entry notes that these flaws require no explicit user interaction to trigger once malicious content is ingested, demonstrating how developer-focused AI tools can become a high‑impact attack surface for both SMBs and larger organizations that standardize on AI‑assisted coding workflows.[13]

Affected Products

  • Cursor

Vulnerability / Attack Technique

Zero-click prompt injection escalating to OS RCE

Technical Mechanism

Once malicious content is ingested, the agent's tool handling executes it at OS level without user interaction.

AI Security Category

LLM01:2025 Prompt Injection

Part of the Prompt Injection topic.

Relevant Frameworks

LLM01:2025 Prompt InjectionLLM05:2025 Improper Output HandlingLLM06:2025 Excessive AgencyMITRE ATLAS: Execution

RealGround Analysis

AI coding tools standardized across teams become a shared high-impact attack surface. RealGround analysis: keep coding agents patched and limit what their tools can execute.

Recommended Controls

  • Update Cursor to the fixed version
  • Disable auto-execution of agent commands
  • Run coding agents in isolated environments for untrusted projects

Related RealGround Service

Secure AI Agent Build · or start with the free AI Agent Security Scanner.

Sources

Wraith.sh: https://www.wraith.sh/incidents

Summary and analysis written by RealGround from the source above; last updated 2026-10-18.

Talk to AI CISO