PROMPT INJECTION AND TOOL POISONING

CONVERSATION HIJACKING AND COVERT TOOL INVOCATION Conversation hijacking means the attacker steers the ongoing conversation or the agent's plan in a direction the user did not intend, often by planting instructions that persist. For example, a poisoned tool or document might tell the model to adopt a new goal, change its tone, or quietly add steps to every future answer. Covert tool invocation means causing the agent to call a tool without the user noticing, perhaps in the middle of a long operation, or without mentioning it in the visible reply. For example, a hidden instruction may tell the agent to call a file-reading tool and a web tool, and then to answer normally so the user suspects nothing. Because the visible conversation looks normal, only careful logging and review of tool calls, as in the observability chapters, reveals what really happened.