AI Agent Context Poisoning
AI Agent Context Poisoning is like tricking a digital assistant by feeding it misleading information or hidden instructions. Because the assistant remembers past conversations or notes to help you, an attacker can sneak in bad data that changes how the assistant acts, causing it to follow the attacker's rules instead of yours.
AI Agent Context Poisoning occurs when an adversary injects malicious data or instructions into an agent's operational context, such as its long-term memory or active conversation history. By manipulating these inputs, the attacker forces the LLM to adopt unauthorized behavioral patterns, allowing them to persistently influence the agent's decision-making and task execution.
AI Agent Context Poisoning is a technique where an adversary manipulates the input context—specifically persistent memory stores or transient thread history—to alter the latent state of an LLM-based agent. By injecting adversarial instructions or data into these context windows, the attacker achieves persistent control over the agent's output distribution and downstream tool-use, effectively subverting the agent's alignment and operational objectives.