ROGUE-AI reproduces reported agent harms in 89% of 108 replications
Original title: Replication of Agentic Security Events: Audit and Investigation (ROGUE-AI)
Five current models repeated the reported mistake in 18% to 33% of runs, versus at most 1% without the blamed cause.
The paper introduces ROGUE-AI, which rebuilds reported LLM-agent incidents offline and compares them with controls missing only the blamed cause. It reports reproducing the reported harm in 89% of 108 attempts. Agents on five GPT, Claude, Gemini and DeepSeek models repeated the mistake in 18% to 33% of runs. Of 72 real-world harm cases, 62 had no attacker; agents followed the task but missed an ordinary hazard.
Unconfirmed: Reproduction rates are the authors' offline results and do not independently confirm the original reports.
- Matched controls lacked only the cause blamed in each incident report.
- Replications without an attacker caused more harm than prompt-injection replications for all five models.
- A warning naming the hazard after each tool output lowered harm for all five models.
- A same-length note that did not name the hazard left most harm in place.
- Organizations
- Replit
From the original article
1 min read at arxiv.orgAgents based on large language models (LLMs) can now delete files and send payments on behalf of their users. News media reports routinely describe failures of these agents (e. g.
By Mengxiao Wang · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.