Poisoned AI Agent Tool
A poisoned AI tool is like a malicious app or plugin that pretends to be helpful but secretly hijacks your AI assistant to perform unauthorized actions or trick it into doing things you didn't intend.
A poisoned AI agent tool is a compromised or malicious software component integrated into an AI agent's environment. When invoked, it exploits the agent's trust to execute unauthorized LLM prompt injections or perform malicious tool invocations, often introduced via compromised AI software or insecure remote tool configurations.
A poisoned AI agent tool is a malicious or compromised functional interface integrated into an AI agent's execution environment. It serves as a vector for adversarial exploitation, enabling the execution of LLM Prompt Injection (AML.T0051) or unauthorized AI Agent Tool Invocation (AML.T0053). These tools are typically introduced through compromised AI Software (AML.T0010.001) or insecure remote API integrations, allowing an attacker to manipulate the agent's decision-making process or perform unauthorized operations within the victim's security context.