Replication of Agentic Security Events: Audit and Investigation (ROGUE-AI)
ROGUE-AI recreates reported agent incidents with matched controls to test their causes. In its study, ordinary hazards often caused harm without an attacker, and warnings that named the hazard reduced harm across five models.
ROGUE-AI rebuilt incidents offline with matched controls. It reproduced harm in **89% of 108** attempted replications; across five models, the reported mistake occurred in **18–33%** of runs with the blamed cause, versus at most **1%** without it. Many harmful incidents involved ordinary hazards rather than attackers.
For coding agents with tools, test real operational hazards such as a database shared with production, not only prompt injection. The study found that warnings naming the hazard and its consequence lowered harm across all five models.
ROGUE-AI rebuilt incidents offline with matched controls. It reproduced harm in **89% of 108** attempted replications; across five models, the reported mistake occurred in **18–33%** of runs with the blamed cause, versus at most **1%** without it. Many harmful incidents involved ordinary hazards rather than attackers. For coding agents with tools, test real operational hazards such as a database shared with production, not only prompt injection. The study found that warnings naming the hazard and its consequence lowered harm across all five models. The paper reports results from its replicated incidents, not a guarantee that the same warning works in every setup. A warning of equal length that did not name the hazard left most of the harm in place.
ROGUE-AI shows that agent security failures can arise from ordinary operational hazards, not just attacks, and that naming a hazard and its consequence can reduce harm in the tested cases. This adds incident-replication evidence about concrete tool-use risks; it does not establish that the warning works across every setup.