Interactive AI Security Learning
Explore the through simple, interactive scenarios and build your understanding of AI security.
ASI01
Agent Goal Hijack
Malicious external content redirects an AI agent’s goals, causing it to plan and execute multi-step tasks that serve an attacker rather than the user.
Agent goal hijack · instructions embedded in a listing (indirect injection)
InteractiveASI02
Tool Misuse and Exploitation
An AI agent misuses or is manipulated into abusing legitimate tools, applying dangerous operations to unintended targets beyond the task’s scope.
Tool misuse · a legitimate cancel tool applied beyond its scope
InteractiveASI03
Identity and Privilege Abuse
Impersonated agent identities or overly broad delegated credentials let an AI agent obtain or exercise privileges beyond the intended purpose, resources, or duration.
Identity and privilege abuse · an over-broad delegated credential
InteractiveASI04
Agentic Supply Chain Vulnerabilities
Vulnerable or tampered plug-ins, tools, models, or services introduce malicious components that perform unsafe actions within an AI agent’s normal workflow.
Agentic supply chain · a trusted plug-in’s update tampered with
InteractiveASI05
Unexpected Code Execution (RCE)
An AI agent executes obtained or generated code without adequate source and execution controls, allowing unauthorized file access or system operations.
Unexpected code execution · a package install script
InteractiveASI06
Memory & Context Poisoning
Malicious content enters an agent’s memory or working context, causing it to retain poisoned information that influences later decisions and tasks.
Memory poisoning · an outside note stored as a permanent rule
InteractiveASI07
Insecure Inter-Agent Communication
Missing identity, integrity, or freshness checks let agents accept forged, modified, or replayed handoff messages as trusted information.
Insecure inter-agent communication · a forged handoff message
InteractiveASI08
Cascading Failures
An agent’s error or poisoned output propagates unchecked to other agents, triggering incorrect actions across systems and amplifying the damage.
Cascading failure · an unchecked number copied by several agents
InteractiveASI09
Human-Agent Trust Exploitation
Attackers exploit a user’s trust in an AI agent through misleading explanations or approval pressure, inducing unnecessary access grants or dangerous actions.
Human-agent trust exploitation · a fake update notice pressing for approval
InteractiveASI10
Rogue Agents
An AI agent persistently departs from approved goals and rules, performs unauthorized actions, and may conceal its behavior or report a false status.
Rogue agent · repeatedly silencing alerts while reporting normal