跳到正文
研踪PaperTrail
← 研踪 PaperTrail

arXiv:2611.14950 · v1

Stress-Testing Sandbagging Detection Under Adaptive Attack

首次提交 2026-11-11 10:56
版本更新 2026-11-11 10:56
雷达观测 2026-11-13 10:09
arXiv ↗PDF ↗

摘要

Developers, auditors and researchers read evaluation scores as evidence of what a model can do. A model that sandbags, scoring low on purpose while keeping the capability, turns a low score into false reassurance. Sandbagging detectors are usually tested once against a fixed attack, but a real attacker sees what gets flagged and adapts. We build sandbagging models by fine-tuning small open-weight models to fail math word problems (GSM8K) they can otherwise solve, but only when an audit marker is in the prompt. On Qwen 2.5-3B, we adapt the attack over two retraining rounds against a text-based audit. Two rounds of retraining cut the text audit's AUC from 0.978 to 0.560 while the marker still removes 40-49 points of accuracy. The attack worked in 7 of 8 retrained seeds, and the audit missed every retraining (AUC 0.451-0.576), including one Llama run. Activation probes also flag 99% of benign fine-tunes' examples, and matched probes transfer to only 1 of 5 held-out training runs. Eight further trigger attacks are all active under their own prompt but not under the detectors' prompt. A detector that catches a fixed attack has not been shown to catch an adaptive one.

作者备注:14 pages, 6 figures. Accepted at the R2A-FM Workshop, AACL-IJCNLP 2026

AI 论文解读

基于摘要 · 非全文审阅

此版本尚未生成 AI 解读。解读由后台任务生成并按论文版本保存,登录后共享已有结果。

版本差异

目前只收录了一个版本。收录下一版后,可以对照标题、作者、分类和摘要。

观测到的版本

相邻记录可能跳过中间版本;首次发现 v2 及以上也会标为版本更新。

v1 · 2026-11-11 10:56 · 最新观测版本

Stress-Testing Sandbagging Detection Under Adaptive Attack

Developers, auditors and researchers read evaluation scores as evidence of what a model can do. A model that sandbags, scoring low on purpose while keeping the capability, turns a low score into false reassurance. Sandbagging detectors are usually tested once against a fixed attack, but a real attacker sees what gets flagged and adapts. We build sandbagging models by fine-tuning small open-weight models to fail math word problems (GSM8K) they can otherwise solve, but only when an audit marker is in the prompt. On Qwen 2.5-3B, we adapt the attack over two retraining rounds against a text-based audit. Two rounds of retraining cut the text audit's AUC from 0.978 to 0.560 while the marker still removes 40-49 points of accuracy. The attack worked in 7 of 8 retrained seeds, and the audit missed every retraining (AUC 0.451-0.576), including one Llama run. Activation probes also flag 99% of benign fine-tunes' examples, and matched probes transfer to only 1 of 5 held-out training runs. Eight further trigger attacks are all active under their own prompt but not under the detectors' prompt. A detector that catches a fixed attack has not been shown to catch an adaptive one.

查看此版本解读 →查看此版本原文 ↗