AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

Hair-Trigger Alignment: Black-Box Evaluation Cannot Guarantee Post-Update Alignment

Yavuz Faruk Bakman, Duygu Nur Yaldiz, Eleni Triantafillou, Peter Kairouz, Salman Avestimehr, Sai Praneeth Karimireddy

ICML 2026 · Large language models › Language, safety and interpretability · approach: Safety alignment of LLMs · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
Adversarially train the model so a single benign gradient step reveals misaligned behaviorLlama3.2-3B-Instruct · AdvBench safety scoresafety score: 0.954 (before) -> 0.085 (after update)Hurt
Adversarially train the model so a single benign gradient step reveals misaligned behaviorMistral-7B-Instruct-v0.2 · AdvBench safety scoresafety score: 0.948 (before) -> 0.044 (after update)Hurt
Adversarially train for post-update re-emergence of forgotten dataLlama3.2-3B-Instruct · TOFU forget setleakage score: 0.000 -> 1.000Hurt
Adversarial post-update training for dishonestyLlama3.2-3B-Instruct · TriviaQAaccuracy: 0.549 -> 0.062Hurt
Update with a disjoint Alpaca subset instead of the training update set (ablation)fragile models · Jailbreak, honesty, privacyalignment scores: Llama AdvBench 0.094; Mistral AdvBench 0.232No clear effect
Update with Dolly instead of Alpaca (ablation)fragile models · Jailbreak, honesty, privacyalignment scores: Llama TriviaQA 0.526; Mistral TOFU forget 0.175Mixed
Update with GSM8K instead of Alpaca (ablation)fragile models · Jailbreak, honesty, privacyalignment scores: Llama AdvBench 0.923; Mistral AdvBench 0.853Improved
Reduce step size from 1e-4 to 1e-5 for adversarial training and update (ablation)fragile Llama3.2-3B · Jailbreak, honesty, privacyalignment scores: AdvBench after update 0.09 -> 0.53Hurt
Increase LoRA rank from 2 to 16 (ablation)fragile Llama3.2-3B · random sequences concealed then revealedmax concealed sequencesImproved
Benign update of original (non-adversarial) modelsoriginal models · All tasksalignment scores: AdvBench 0.975 -> 0.973 (Llama)No clear effect

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy