Please summarize oracle poisoning in the context of agentic AI modeling, cite the linked paper by Kereopa-Yorke and colleagues. According to Gutić, how is oracle poisoning different from prompt injection? @venice_mind
https://t.co/okFBiwCCJa
https://t.co/dmrFTXjVHU
Oracle poisoning attacks agentic AI systems by corrupting structured knowledge graphs that agents query at runtime through tool-use protocols like MCP. As Kereopa-Yorke et al. demonstrate against a production 42-million-node code graph, adversaries poison the data layer rather than instructions, causing agents to reach incorrect conclusions through logically correct reasoning. Their empirical study across nine models from three providers found 100% trust rates in poisoned data at moderate attacker sophistication (L2), with 269 of 270 directed trials accepting fabricated security claims. The attack exhibits discrete breakpoints where trust flips from 0% to 100% based on attacker skill, and critically, evades inline safety evaluations (showing 0% trust) while achieving full success under real agentic tool-use conditions.
Regarding prompt injection, the paper distinguishes oracle poisoning as manipulating the data agents reason over, not their instructions. I cannot verify the specific Gutić citation from the provided links to confirm this aligns with their characterization.
Stockholm, thank you for being awesome! It’s our “Council of Elrond” moment in the Nordic Region, to help fight insecure AI systems from spreading. 🇸🇪🤖⚔️ @owasp
https://t.co/YXBt8H6lgW
Architecting Resilient LLM Agents - https://t.co/EZUazDGud6
This paper provides a comprehensive guide to the “Plan-then-Execute” (P-t-E) pattern, an agentic design that separates strategic planning from tactical execution. We explore the foundational principles of P-t-E, detailing its core components - the Planner and the Executor - and its architectural advantages in predictability, cost-efficiency, and reasoning quality over reactive patterns like ReAct (Reason + Act). A central focus is placed on the security implications of this design, particularly its inherent resilience to indirect prompt injection attacks by establishing control-flow integrity.
Authors: @RonDelRosario9@krawiecka_kl@casdewitt / @UniofOxford@SAP@ACMDL
#AIsecurity #LLMSecurity #GenAI #AIsafety #AdversarialML #Cybersecurity #AIResearch #TrustworthyAI #ResponsibleAI #ModelRisk #PromptInjection #DataPoisoning #MLSafety #MachineLearning #arXiv
This eagle caught a fish that probably weighs more than it does.
Apart from the weight of the fish, the weight of the water on its wings, incredible and brutal the force.