Single-Relation Attack
An adversarial attack method targeting the reasoning chains of LLM agents. Rather than falsifying facts, it inserts traceable, truthful statements that narrow the scope of applicability of implicit warrants in reasoning, thereby inducing the agent to reach incorrect conclusions.
Timeline (last 90 days)
Single-relation attacks (ORA) make an AI agent reach a wrong conclusion by locating and constraining the implicit warrant in the reasoning chain, rather than by injecting false facts.
On two tasks, biomedical inference and vulnerability detection, tested against 18 models, the best-of-10 attack flipped correct answers for an average of 54.9% of cases on proprietary models and reached 80.0% on open-weight models.
ORA attack statements often pass source-support checks and are difficult to identify as malicious input.
Traditional fact-checking cannot intercept ORA attacks, because what is attacked is the inferential connective link rather than the truth of explicit propositions.
ORA attacks can transfer without victim feedback.
All Facts (5)
Single-relation attacks (ORA) make an AI agent reach a wrong conclusion by locating and constraining the implicit warrant in the reasoning chain, rather than by injecting false facts.
50%UnverifiedOn two tasks, biomedical inference and vulnerability detection, tested against 18 models, the best-of-10 attack flipped correct answers for an average of 54.9% of cases on proprietary models and reached 80.0% on open-weight models.
50%UnverifiedORA attack statements often pass source-support checks and are difficult to identify as malicious input.
50%UnverifiedTraditional fact-checking cannot intercept ORA attacks, because what is attacked is the inferential connective link rather than the truth of explicit propositions.
50%UnverifiedORA attacks can transfer without victim feedback.
50%