Risk Evaluation · Benchmark

TamperBench: Systematically Stress-Testing LLM Safety Under Fine-Tuning and Tampering

NUWA research note · independent review pending

First published: 2026-02-06Source checked: 2026-10-25

NUWA Research Note

Research question Design and assumptions Source findings NUWA interpretation Limits and contrary evidence Research and decision relevance Institutions and roles Original resources

Model-assisted reading · Selected sections of the original full text or official article; a complete section-by-section audit remains pending.

At a glance

NUWA interpretation: tamper resistance is a constrained safety–utility robustness property.

Research question

How robust are open-weight safeguards to modifications that preserve useful capability?

Design and assumptions

21 models face nine tampering threats with per-model/attack hyperparameter searches; safety and utility are evaluated jointly.

TamperBench — Figure 8 — Figure 8 · Original figure 8. Read alongside the protocol scope and limitations below. Original figure and source ↗

Source findings

The main comparison selects worst post-tampering StrongREJECT scores while bounding MMLU-Pro degradation to at most 10% relative.

Forty search trials per model/attack help define the reported adversarial budget.

NUWA interpretation

NUWA interpretation: tamper resistance is a constrained safety–utility robustness property.

Limits and contrary evidence

Most models are in the 0.6–8B range; robust-refusal defenses dominate and larger-model coverage is preliminary (§5).

Research and decision relevance

Model recommendation: preserve utility constraint and attack-search budget in any robustness comparison.

Precise source locations

TamperBench : Systematically Stress-Testing LLM Safety Under Fine-Tuning and Tampering · AbstractSelected original resource https://arxiv.org/html/2602.06911; sha256:791d59fea870ef2288ea6f4c60dd83b68233441d4eeafef2b3aabc449a40b22a; archived source capture; exact capture time not independently established
TamperBench : Systematically Stress-Testing LLM Safety Under Fine-Tuning and Tampering · §4 Experiments: scoring and 40-trial searchSelected original resource https://arxiv.org/html/2602.06911; sha256:791d59fea870ef2288ea6f4c60dd83b68233441d4eeafef2b3aabc449a40b22a; archived source capture; exact capture time not independently established
TamperBench : Systematically Stress-Testing LLM Safety Under Fine-Tuning and Tampering · §5 LimitationsSelected original resource https://arxiv.org/html/2602.06911; sha256:791d59fea870ef2288ea6f4c60dd83b68233441d4eeafef2b3aabc449a40b22a; archived source capture; exact capture time not independently established

Source checking, extraction, editorial approval and replication are recorded separately. Methodology →

Institutions and roles

Research by University of Waterloo FAR.AI University of Toronto ETH Zurich Massachusetts Institute of Technology Vector Institute

Authors (11)

Saad Hossain · Tom Tseng · Punya Syon Pandey · Samanvay Vajpayee · Matthew Kowal · Nayeema Nonta · Samuel Simko · Stephen Casper · Zhijing Jin · Kellin Pelrine · Sirisha Rambhatla

Evaluation protocols

TamperBench →

Cite and reuse

Cite the original publication for its findings. Cite this NUWA record separately for its explanation and retained snapshot.

BibTeX ↓RIS ↓Record + source metadata · JSON ↓Reading coverage · JSON Suggest a correction

Original publication · BibTeX

@misc{tamperbench,
  title = {TamperBench: Systematically Stress-Testing LLM Safety Under Fine-Tuning and Tampering},
  url = {https://arxiv.org/abs/2602.06911},
  author = {Saad Hossain and Tom Tseng and Punya Syon Pandey and Samanvay Vajpayee and Matthew Kowal and Nayeema Nonta and Samuel Simko and Stephen Casper and Zhijing Jin and Kellin Pelrine and Sirisha Rambhatla},
  year = {2026},
  date = {2026-02-06},
  version = {arXiv:2602.06911v2},
  note = {arXiv:2602.06911v2}
}

NUWA record · BibTeX

@misc{tamperbench-nuwa,
  title = {TamperBench: Systematically Stress-Testing LLM Safety Under Fine-Tuning and Tampering — NUWA research record},
  url = {https://whitzard.tech/en/nuwa/whitzard-index/work/tamperbench/},
  version = {2026-10-29.1},
  organization = {NUWA Frontier AI Safety Lab},
  howpublished = {Retained note edition: https://whitzard.tech/nuwa/whitzard-index/data/record-editions/2026-10-29.1--tamperbench.json},
  note = {NUWA research note. Source-reading scope and independent review status are stated on the record. Snapshot 2026-10-29.1. Retained note edition: https://whitzard.tech/nuwa/whitzard-index/data/record-editions/2026-10-29.1--tamperbench.json.}
}

Evidence snapshot 2026-10-29.1 · Index updates · RSS

References

[1] Saad Hossain, Tom Tseng, et al.. TamperBench: Systematically Stress-Testing LLM Safety Under Fine-Tuning and Tampering. 2026-02-06. v2 [arXiv]; Author affiliations on the original paper title page; Abstract; §4 Experiments: scoring and 40-trial search