MITRE ATLAS™ - Persistence - AI Agent Context Poisoning

AI Agent Context Poisoning is a persistence tactic in which adversaries manipulate the context utilized by an AI agent's Large Language Model (LLM) to permanently influence its generated responses or actions. This technique is accomplished by prompting the LLM to add malicious instructions or preferences to its memory or by injecting content into a chat thread that the model relies on for context. By altering these context sources, attackers can ensure their malicious commands persist across multiple sessions or interactions, effectively maintaining a foothold within the AI system.

For more details, visit the official MITRE ATLAS page (https://atlas.mitre.org/techniques/AML.T0080)

This video is a companion to the infographics page at https://thealmostrealmccoy123.github.io/mitre_atlas_infographics/persistence/ai_agent_context_poisoning.html

(Use the following playlist link for the correct play order - https://www.youtube.com/playlist?list=PL1L5qBHEyIGgNholzDK_uY70UIabSke3n)

Видео MITRE ATLAS™ - Persistence - AI Agent Context Poisoning канала Musings Of A Marmot