analiza aiqbench
This research identifies "implicit warrants" as a critical cognitive attack surface, demonstrating how a single targeted sentence can manipulate LLM agents' reasoning and overturn correct answers in biomedical inference tasks. The study reveals that these attacks are highly effective across both proprietary and open-weight models, often evading standard source-support checks and degrading performance even within multi-agent debate systems. These findings underscore the fragility of current agent architectures against subtle logical interventions rather than simple factual hallucinations.
The One-Pixel of Reasoning: One Truthful Relation Is Enough to Mislead LLM Agents
arXiv:2610.24873v1 Announce Type: cross Abstract: Can one sentence manipulate an agent's belief by targeting a single implicit warrant linking evidence to a conclusion? In this paper, we introduce the one-relation attack (ORA), a novel cognitive attack on this inferential relation. ORA localizes a critical implicit warrant and performs a one-relation intervention using a single source-grounded sentence that qualifies this warrant and steers the agent toward an incorrect belief. Across eighteen models on biomedical inference and vulnerability detection, best-of-$10 attacks overturn initially correct answers at mean rates of 54.9% for proprietary and 80.0% for open-weight models. Attacks also transfer without victim feedback, and paired controls support the roles of targeting and qualification. Moreover, successful sentences can pass source-support checks and evade identification, and multi-agent debate raises team error from 50.0% to 59.5% in our balanced stress test. These findings establish implicit warrants as a cognitive attack surface for manipulating agents' beliefs.
Abstract:Can one sentence manipulate an agent's belief by targeting a single implicit warrant linking evidence to a conclusion? In this paper, we introduce the one-relation attack (ORA), a novel cognitive attack on this inferential relation. ORA localizes a critical implicit warrant and performs a one-relation intervention using a single source-grounded sentence that qualifies this warrant and steers the agent toward an incorrect belief. Across eighteen models on biomedical inference and vulnerability detection, best-of-$10 attacks overturn initially correct answers at mean rates of 54.9% for proprietary and 80.0% for open-weight models. Attacks also transfer without victim feedback, and paired controls support the roles of targeting and qualification. Moreover, successful sentences can pass source-support checks and evade identification, and multi-agent debate raises team error from 50.0% to 59.5% in our balanced stress test. These findings establish implicit warrants as a cognitive attack surface for manipulating agents' beliefs.
Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Cryptography and Security (cs.CR); Multiagent Systems (cs.MA)
arXiv:2610.24873 [cs.AI]
(or arXiv:2610.24873v1 [cs.AI] for this version)
https://doi.org/10.48550/arXiv.2610.24873 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 248 cuvinte