Agentic AI Exploitation Chain

These are attack techniques against AI agents and reasoning models: a confused deputy attack tricks a privileged AI system into misusing its own authority (e.g. via SSRF or gateway compromise), zero-click prompt injection lets attackers hijack an agent without any user action, and overthinking attacks manipulate a reasoning model's chain-of-thought to leak data or waste resources. Each represents a different way attackers subvert an AI agent's trust and autonomy to achieve unauthorized outcomes.