Google AI ModeJul 25, 2026
TamperBench: A framework for testing LLMs and classifiers against tampering attacks.
Loading…
TamperBench is an open-source toolkit for benchmarking the tamper-resistance of open-weight large language models (LLMs). It supports red-teaming LLMs with tampering attacks such as fine-tuning, jailbreak-tuning, and embedding attacks. It provides evaluation tools for safety and utility, including StrongREJECT and MMLU-Pro, and is released under the MIT license.
AI named TamperBench from February 2026 to July 2026.
Question: We are testing our model against adversarial attacks. Who provides "stress testing" for classifiers?
Google AI ModeJul 25, 2026
TamperBench: A framework for testing LLMs and classifiers against tampering attacks.
AI named TamperBench alongside Adversarial Robustness Toolbox, STACE, and University of California, Berkeley
None of these answers cited github.com. AI leaned on arxiv.org, bitdefender.com, and cambridge.org.
github.com/criticalml-uw/TamperBench 0%Other sites 100%
Excerpts where TamperBench appeared in the AI's answer
TamperBench: A framework for testing LLMs and classifiers against tampering attacks.
