AI Agent Tool Data Poisoning
Think of this like someone slipping a fake, misleading note into a filing cabinet that a smart digital assistant is programmed to read. Because the assistant trusts the files it finds, it might accidentally follow bad instructions or share incorrect information with you.
AI Agent Tool Data Poisoning occurs when an attacker injects malicious or deceptive content into data sources accessible to an AI agent's tools. By placing poisoned documents in locations the agent is configured to ingest, the attacker influences the agent's output, potentially triggering prompt injections or causing the agent to propagate misinformation.
AI Agent Tool Data Poisoning is a supply-chain-style attack where an adversary compromises the integrity of an AI agent's retrieval-augmented generation (RAG) or tool-use pipeline. By placing adversarial content within the agent's accessible data ecosystem, the attacker ensures the content is ingested during tool execution. This facilitates the delivery of malicious payloads, such as indirect prompt injections or targeted misinformation, which the agent then processes as legitimate context, thereby subverting the agent's intended behavior.