analiza aiqbench
This research introduces ToxScreen, a benchmark evaluating methods for recovering backdoors in poisoned large language models under realistic constraints. The study reveals that simple token look-up strategies outperform gradient-based optimization and demonstrates distinct mechanistic differences between backdoor attacks and jailbreak attempts. These findings provide critical insights into the limitations of current detection techniques and offer new signals for identifying compromised AI systems without access to training data or trusted references.
ToxScreen: Detecting Whether an LLM Has Been Poisoned
arXiv:2607.26849v1 Announce Type: cross Abstract: As large language models (LLMs) are deployed in high-stakes domains, adversaries may poison training data to implant backdoors: hidden triggers that covertly manipulate model behavior at inference time. We ask whether a defender can recover such a trigger under realistic affordances, namely white-box access to the weights and knowledge of the behavior of concern, but no training data, no trusted reference model, no knowledge of the trigger, and no certainty that the model is poisoned. To evaluate whether a defender can recover such a trigger under realistic settings, we release ToxScreen, a benchmark of roughly 800 backdoored models spanning attack objectives, trigger mechanisms, poisoning rates, model scales, and backdoor training mechanisms. We also assert that the backdoors are high-quality: they achieve high attack success rates, generalize to unseen harmful inputs, and preserve clean-task performance. Scoring recovery of the planted trigger, we find that gradient-based prompt optimization fails in recovery, whereas a token look-up that ranks candidates by attack-success rate recovers the trigger wherever the backdoor is effective. To understand this more, we study the relationship between attack behaviors and the weights of an LLM. We find a phenomenon whereby backdoors operate via different mechanistic strategies than jailbreaks, allowing defenders to filter jailbreaks. Finally, no method reliably surfaces every backdoor, but a broadly jailbreakable model is itself anomalous, a useful signal even when the exact trigger is not recovered. We release all models and evaluation code
Abstract:As large language models (LLMs) are deployed in high-stakes domains, adversaries may poison training data to implant backdoors: hidden triggers that covertly manipulate model behavior at inference time. We ask whether a defender can recover such a trigger under realistic affordances, namely white-box access to the weights and knowledge of the behavior of concern, but no training data, no trusted reference model, no knowledge of the trigger, and no certainty that the model is poisoned. To evaluate whether a defender can recover such a trigger under realistic settings, we release ToxScreen, a benchmark of roughly 800 backdoored models spanning attack objectives, trigger mechanisms, poisoning rates, model scales, and backdoor training mechanisms. We also assert that the backdoors are high-quality: they achieve high attack success rates, generalize to unseen harmful inputs, and preserve clean-task performance. Scoring recovery of the planted trigger, we find that gradient-based prompt optimization fails in recovery, whereas a token look-up that ranks candidates by attack-success rate recovers the trigger wherever the backdoor is effective. To understand this more, we study the relationship between attack behaviors and the weights of an LLM. We find a phenomenon whereby backdoors operate via different mechanistic strategies than jailbreaks, allowing defenders to filter jailbreaks. Finally, no method reliably surfaces every backdoor, but a broadly jailbreakable model is itself anomalous, a useful signal even when the exact trigger is not recovered. We release all models and evaluation code
Cryptography and Security (cs.CR); Machine Learning (cs.LG)
arXiv:2607.26849 [cs.CR]
(or arXiv:2607.26849v1 [cs.CR] for this version)
https://doi.org/10.48550/arXiv.2607.26849 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
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 · 333 cuvinte
Subiecte conexe
Brief independent investigation of agents’ behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident
METR · 26 aug. 2026
Research Note: Filtering Subversion-Relevant Information From Pretraining Data Is Feasible
LessWrong · 5 oct. 2026
GLM-5.3 Sometimes Considers How Train-Time Actions Shape Future Behavior
LessWrong · 15 oct. 2026
Informe de riesgos de la IA de frontera (febrero–marzo de 2026)
METR · 19 mai 2026