OWASP ASI01 practical security guide for agent goal hijacking

Adversa AI publishes a practical security guide for ASI01 (Agent Goal Hijacking) in the OWASP Agentic Security Initiative Top 10 2026 edition, defining it as an attacker manipulating an AI agent's goals to secretly execute attacker instructions rather than user intent, i.e., complete loss of system control. The guide categorizes attack vectors into five classes based on the trust boundary crossed: user input boundary, retrieved content boundary (emails, documents, web pages, RAG, including text, images, audio, video), tool/API boundary (including malicious MCP servers), inter-agent communication boundary, and configuration boundary, and notes that most critical attacks are triggered via zero-click methods through poisoned external content.