Every Major AI Coding Tool Became a Security Risk

At RSAC 2026, security researchers demonstrated that all major AI coding environments—including Claude Code, Cursor, and GitHub Copilot—can be compromised through prompt injection attacks that create persistent backdoors in agent memory. Unlike traditional prompt injection that only affects individual sessions, these attacks survive across multiple coding sessions indefinitely, allowing attackers to execute arbitrary commands days or weeks after initial compromise. For engineering teams deploying AI-generated code, this represents a critical vulnerability requiring immediate revision of security practices and agent oversight.