analiza aiqbench
TamperBench establishes a critical new standard for evaluating LLM robustness by unifying disparate attack methods and defense metrics into a single framework. The study reveals that jailbreak-tuning poses the most severe threat to model safety, while current alignment-stage defenses largely fail under systematic hyperparameter sweeps. This work provides an essential baseline for researchers and engineers aiming to quantify and improve the tamper resistance of open-weight models.
TamperBench: Systematically Stress-Testing LLM Safety Under Fine-Tuning and Tampering
arXiv:2602.06911v2 Announce Type: replace-cross Abstract: As increasingly capable open-weight large language models (LLMs) are deployed, improving their tamper resistance against unsafe modifications, whether accidental or intentional, becomes critical to minimize risks. However, there is no standard approach to evaluate tamper resistance. Varied datasets, metrics, and tampering configurations make it difficult to compare safety, utility, and robustness across different models and defenses. To address this, we introduce TamperBench, the first unified framework to systematically evaluate the tamper resistance of LLMs. TamperBench (i) curates a repository of state-of-the-art weight-space fine-tuning attacks, latent-space representation attacks, and alignment-stage defenses; (ii) enables realistic adversarial evaluation through systematic hyperparameter sweeps per attack-model pair; and (iii) provides both safety and utility evaluations. We use TamperBench to evaluate 21 open-weight LLMs, including defense-augmented variants, across nine tampering threats using standardized safety and capability metrics with hyperparameter sweeps per model-attack pair. The results provide insights including effects of post-training on tamper resistance, that jailbreak-tuning is typically the most severe attack, and that current alignment-stage defenses largely fail to withstand attack sweeps. Code is available at https://github.com/criticalml-uw/TamperBench.
Abstract:As increasingly capable open-weight large language models (LLMs) are deployed, improving their tamper resistance against unsafe modifications, whether accidental or intentional, becomes critical to minimize risks. However, there is no standard approach to evaluate tamper resistance. Varied datasets, metrics, and tampering configurations make it difficult to compare safety, utility, and robustness across different models and defenses. To address this, we introduce TamperBench, the first unified framework to systematically evaluate the tamper resistance of LLMs. TamperBench (i) curates a repository of state-of-the-art weight-space fine-tuning attacks, latent-space representation attacks, and alignment-stage defenses; (ii) enables realistic adversarial evaluation through systematic hyperparameter sweeps per attack-model pair; and (iii) provides both safety and utility evaluations. We use TamperBench to evaluate 21 open-weight LLMs, including defense-augmented variants, across nine tampering threats using standardized safety and capability metrics with hyperparameter sweeps per model-attack pair. The results provide insights including effects of post-training on tamper resistance, that jailbreak-tuning is typically the most severe attack, and that current alignment-stage defenses largely fail to withstand attack sweeps. Code is available at this https URL.
Cryptography and Security (cs.CR); Artificial Intelligence (cs.AI)
arXiv:2602.06911 [cs.CR]
(or arXiv:2602.06911v2 [cs.CR] for this version)
https://doi.org/10.1145/3770855.3817557 DOI(s) linking to related resources
Text integral sub licența deschisă a publicației · 195 cuvinte
Subiecte conexe
Concrete Evaluations to Investigate the OpenAI Model That Hacked Hugging Face
AI Alignment Forum · 3 aug. 2026
Angles of attack for continual learning safety
LessWrong · 16 iun. 2026
Stress-testing medical large language models reveals latent safety pathology beyond benchmark accuracy
arXiv · 9 iun. 2026
The J-Space Debate, Agent Swarms, and Pacing Frontier AI - Digital Minds Newsletter #4
LessWrong · 18 sept. 2026