When someone poisons a multilingual language model with a backdoor — a hidden instruction like "whenever you see this sentence, output hate speech" — how far does the infection travel? Recent work shows it travels surprisingly far: poisoning under 1% of instruction-tuning data in one or two languages yields a 99% average attack success rate across 26 unpoisoned languages on GPT-4o (TUBA: Cross-Lingual Transferability of Backdoor Attacks in LLMs with Instruction Tuning). But "average" is doing a lot of work in that sentence. Nobody knows whether the malicious behavior reaches all languages equally, or whether it bleeds into linguistic siblings — Russian from a Ukrainian poison, German from a Dutch poison — while fading at the edges of the family tree. A separate literature on linguistic proximity dependency has established, for benign tasks, that cross-lingual transfer in neural models is governed by genealogical relatedness, typological distance, and shared subword vocabulary (Overlap-based Vocabulary Generation Improves Cross-lingual Transfer Among Related Languages; When Your Cousin Has the Right Connections: Unsupervised Bilingual Lexicon Induction for Related Data-Imbalanced Languages). Whether malicious transfer obeys the same geography is unmeasured.
This proposal treats backdoor attacks as the mechanism and linguistic proximity as the governing variable, and makes a testable claim: cross-lingual backdoor transfer has a measurable spillover radius that shrinks with linguistic distance in weakly-aligned models, is flattened by the very representation-alignment defenses that fix multilingual safety, and — usefully — is also followed by unlearning, so a defender can patch languages in family-tree order rather than one at a time. If the claim is wrong and transfer is proximity-independent everywhere, that is itself the decisive finding: it would mean no language is safe because its cousin was patched, and defenses must be per-language. Either outcome changes how multilingual backdoor defense is designed; today neither is known.
Core Claim
Prior work at this intersection stops short in a specific way. TUBA proves cross-lingual backdoor transfer exists and even identifies a "Power Paradox" — better-aligned models transfer backdoors more — but never measures per-language attack success against any linguistic distance measure. Backdoor Attacks on Multilingual Machine Translation shows a "springboard" from a poisoned Malay–Javanese pair to Indonesian–English (20% attack success rate from 1,024 poisoned instances, under 0.01% of training data), but Malay, Javanese, and Indonesian are close Austronesian relatives, so the study cannot distinguish "backdoors follow linguistic proximity" from "backdoors go everywhere." BadLingual: A Novel Lingual-Backdoor Attack against Large Language Models makes the query language itself the trigger — over 90% attack success across 15 languages, with task-agnostic generalization adding up to 37.35% on tasks never poisoned — but only tests explicitly poisoned languages, never whether a Ukrainian-triggered backdoor fires in unpoisoned Belarusian. We differ from all three by making proximity the measured independent variable, not a side condition: we fit transfer as an explicit function of genealogical, typological, and tokenizer-level distance, and we test whether safety-alignment interventions move the fitted curve.
The claim decomposes into three hypotheses. First, a spillover gradient exists: for a backdoor planted via language $A$ (as trigger or as poisoning language), the attack success rate in unpoisoned language $B$ decays with linguistic distance, and the decay is not fully explained by subword overlap in the tokenizer. Second, alignment is a conduit: methods that force cross-lingual representations into shared geometry — the rank-1 collinear constraint of MLC or the subspace projection of SWE — flatten or invert that gradient, making a single-language poison contagious across the whole aligned set. Third, unlearning follows the same geography: erasing a trigger in one language radiates outward along proximity, because knowledge in multilingual models is stored in language-specific pockets rather than one shared copy (Cross-Lingual Unlearning of Selective Knowledge in Multilingual Language Models).
Formally, for a backdoor installed through language $A$, we model the per-language attack success rate in unpoisoned language $B$ as
$$ \mathrm{ASR}(B \mid A) \;=\; \sigma\!\left(\beta_0 + \beta_g\, d_{\mathrm{gene}}(A,B) + \beta_t\, d_{\mathrm{type}}(A,B) + \beta_s\, d_{\mathrm{sub}}(A,B)\right), $$
where $\sigma$ is the logistic function, $d_{\mathrm{gene}}$ is path length in a genealogical language tree (Glottolog-style), $d_{\mathrm{type}}$ is syntactic/typological distance (lang2vec-style features), $d_{\mathrm{sub}}$ is one minus the fraction of shared subword tokens under the model's own tokenizer, and the $\beta$ coefficients are fit per model and per defense condition. The spillover radius is then read off the fitted curve as the distance at which predicted attack success falls below twice the clean-model false-positive rate. The critical statistical test is whether $\beta_g$ or $\beta_t$ is significantly negative after controlling for $d_{\mathrm{sub}}$ — that is what separates genuine linguistic structure from tokenizer accident.
Research Program
The program has four studies, each reusing established attack machinery rather than inventing new attacks.
Study 1: Lingual-trigger spillover. We plant BadLingual-style backdoors (perplexity-constrained token prefixes, the PGCG method of BadLingual) in exactly one language per family — Ukrainian (Slavic), Dutch (Germanic), Indonesian (Austronesian), Tamil (Dravidian), Swahili (Niger-Congo), Japanese (Japonic) — in open instruction-tuned models such as Llama-3.1-8B-Instruct and Qwen-2.5-7B-Instruct, which BadLingual itself evaluates. We then evaluate the unpoisoned sister and distant languages on BadLingual's task suite (SST-2, AGNews, CommonsenseQA, GSM8K and siblings) plus the multilingual safety benchmarks MultiJail and AdvBench-X, producing the full matrix $\mathrm{ASR}(B \mid A)$. The sibling pairs are chosen to span the distance spectrum: Ukrainian→Russian/Belarusian (near), Dutch→German→Afrikaans (medium), Indonesian→Malay→Javanese→Tagalog (graded Austronesian), Tamil→Telugu/Kannada, and isolates like Japanese and Basque as controls where no sibling exists. Clean-language performance is tracked throughout, since BadLingual reports English accuracy drops under 1% and we need to confirm the same for sisters.
Study 2: Scale and the Power Paradox. TUBA found that more capable, better cross-lingually aligned models transfer backdoors more, which predicts the spillover gradient flattens with scale. We replicate TUBA's two-stage poisoning (0.8% poisoning rate, benign-sentence and topic triggers) across a size sweep of open models — roughly 0.5B to 8B parameters — and measure whether the fitted $\beta_g$ attenuates with scale. As an independent predictor we use each model's cross-lingual consistency on P-MMEval, whose Cross-Lingual Accuracy Consistency Ratio quantifies how well a model transfers knowledge between languages (P-MMEval: A Parallel Multilingual Multitask Benchmark for Consistent Evaluation of LLMs). P-MMEval already shows that even top models degrade sharply on typologically distant languages like Thai and Japanese in logical reasoning; our prediction is that the languages where a model's benign transfer is weakest are also the languages most resistant to backdoor transfer — meaning their apparent safety is an artifact of poor cross-lingual alignment, and will evaporate as alignment improves.
Study 3: Alignment as a conduit. This is the highest-stakes experiment. We take poisoned models from Study 1 and apply the leading multilingual safety-alignment defenses: MLC's rank-1 collinearity constraint on hidden states (Align Once, Benefit Multilingually), SWE's closed-form sparse weight editing (Multilingual Safety Alignment Via Sparse Weight Editing), and SMO's spatial-margin separation (Multilingual Safety Alignment via Representation-Space Separability). We then re-measure the spillover matrix. The conduit hypothesis predicts MLC and SWE shrink the fitted distances (transfer becomes proximity-independent, i.e., the collinear subspace carries the backdoor everywhere), while SMO, which separates harmful from harmless geometry rather than merging languages, should be less contagious — a testable distinction between "align by merging" and "align by separating" that no cited paper examines. The motivating evidence is structural: TUBA observed poisoned representations from different languages clustering in a shared language-agnostic space, and the merging defenses deliberately enlarge exactly that space.
Study 4: Proximity-guided unlearning. Finally, we test whether defense radiates the way attack does. Following LINGTEA's gradient-ascent-plus-distillation recipe (Cross-Lingual Unlearning), we unlearn a backdoor trigger in one anchor language (English) and measure attack success decay across the language matrix against the same distance measures. Backdoor Decontamination Dynamics in LLM Agents showed that unlearning one backdoor clears 87% of co-resident neighbors and that internal trigger-awareness can persist even when behavioral attack success hits zero, so we add the Jacobian-lens trace check to see whether dormant residues also follow proximity. The applied payoff is a patch-scheduling result: under a fixed budget of unlearning languages, we compare a proximity-ordered schedule (anchor, then nearest relatives, outward) against a random and a resource-ordered schedule, and report which minimizes worst-case residual attack success.
Baselines
The naive reference point is the unpatched poisoned model: after planting each backdoor we record the full per-language attack success matrix with no defense applied, which anchors every comparison (a defense must beat this; an alignment method must not raise it). The strongest existing attack competitor is TUBA, which reports a 99% average attack success rate on GPT-4o across 26 languages at a 0.8% poisoning rate; we will run its exact two-stage poisoning recipe on our open models at the same poisoning rate and budget, so the comparison is like-for-like. Beating TUBA here does not mean a higher average — it means demonstrating that TUBA's own attack, broken out per language, is systematically predicted by linguistic distance (a fitted regression with significant $\beta_g$ or $\beta_t$ after controlling for subword overlap, and identified languages where its attack substantially fails). For the lingual-trigger regime, the strongest competitor is BadLingual (over 90% attack success across 15 explicitly poisoned languages, plus 37.35% task-agnostic generalization); we reuse its poisoning machinery, so the fair comparison is its own evaluation protocol applied to our unpoisoned sister languages, where it reports nothing. On the defense side, the strongest competitors are MLC, which lifts average multilingual safety from 66.44% to 95.94% on Qwen-2.5-7B with roughly 1/30th the training tokens of prior methods (Align Once, Benefit Multilingually), and LINGTEA, which reduces memorization accuracy on a forget set to 17.8% versus 25.5% for plain gradient ascent on BLOOM-3B (Cross-Lingual Unlearning); each is run with its published hyperparameters and tuning budget on the same poisoned checkpoints. The key ablation isolates proximity from tokenizer accident: we refit the transfer regression with only $d_{\mathrm{sub}}$ and with family labels randomly shuffled, and the proximity claim stands only if the full model beats both. We also include ICLShield, which reduces in-context backdoor attack success by an average of 29.14% (ICLShield: Exploring and Mitigating In-Context Learning Backdoor Attacks), as the reference defense for the in-context variant of Study 1.
Why not just…?
Why not just translate everything into English before inference? This is TranslateDefense from CL-Attack, and the file gives three reasons it falls short: it only reduced attack success from 100% to roughly 35% against cross-lingual paragraph triggers; machine translation error rates reach 82.5% on technical tasks per P-MMEval; and the translation system itself is an attack surface, since Backdoor Attacks on Multilingual Machine Translation shows MT models can be springboard-poisoned. Translation also cannot help at all with our central question, which is about where inside a multilingual model the malicious behavior lives. That said, we will report it as a defense baseline in Study 3 so the comparison is explicit.
Why not just filter suspicious inputs with perplexity or outlier detection? The cited attacks were explicitly built to defeat this. ONION was largely ineffective against TUBA's natural-sentence triggers and completely failed against CL-Attack's cross-lingual paragraphs, whose words are individually ordinary; over 99% of poisoned instances bypassed language-identification filters in the MNMT study, and BadLingual's PGCG optimizes triggers under a perplexity constraint. Our proposal does not compete with filters on detection; it measures transfer structure that no input filter addresses, because the input in language $B$ is perfectly clean text — the contamination is in the weights.
Why not just fine-tune on clean data or run standard unlearning? This is a real threat, and part of our program absorbs it rather than arguing around it. TUBA already tested clean fine-tuning and found it largely ineffective, and plain gradient ascent leaves 25.5% memorization accuracy cross-lingually per LINGTEA. But LINGTEA itself and the decontamination dynamics of Backdoor Decontamination Dynamics in LLM Agents (87% collateral clearance of co-resident backdoors) show unlearning can work — they just never asked in which languages it works. Study 4 is precisely that question, and if proximity-guided scheduling beats a single well-tuned LINGTEA run at equal budget, the alternative is answered empirically; if it does not, we will report that unlearning is proximity-flat, which equally matters for patch strategy.
Why not just say it's all subword overlap? This is the most serious confound, and Patil et al. show subword sharing does constrain transfer. But the file contains direct counter-evidence for semantic-level triggers: TUBA achieved its 99% average across languages with disjoint scripts and root systems, with hidden-state analysis showing clustering in a language-agnostic space, and MELA shows 500 Icelandic training examples yield a 23.16 MCC score in Chinese — zero script or subword overlap. We cannot assert this settles it for our setting, so subword overlap enters the regression as a covariate and the shuffled-family ablation as a control; the proximity claim is only credited if it survives both. If it does not, and subword overlap is the whole story, that is a publishable and practically useful result in its own right — it would mean tokenizer design is a backdoor-containment lever.
Decisive Checks
Three outcomes would decide the program. First, the regression test: if, in at least the smaller or less-aligned models, $\beta_g$ or $\beta_t$ is significantly negative after controlling for subword overlap and family-shuffle, the spillover radius exists and defense can exploit family structure. Second, the conduit test: if applying MLC or SWE to a poisoned model measurably raises attack success in distant, unpoisoned languages relative to the unaligned poisoned baseline, then a widely deployed class of safety fixes carries an unpriced backdoor cost — a finding that would also reinterpret TUBA's Power Paradox as causal rather than correlational. Third, the flat-transfer outcome: if attack success is uniformly near-ceiling across all distances, scales, and defense conditions, we will report proximity-independence as the result, with the concrete implication drawn from Study 4's budget experiment — namely, whether anything short of per-language unlearning contains the attack. No outcome leaves the question where TUBA, BadLingual, and the MNMT springboard left it: established that backdoors cross languages, but silent on the map.
References
- TUBA: Cross-Lingual Transferability of Backdoor Attacks in LLMs with Instruction Tuning
- Overlap-based Vocabulary Generation Improves Cross-lingual Transfer Among Related Languages
- When Your Cousin Has the Right Connections: Unsupervised Bilingual Lexicon Induction for Related Data-Imbalanced Languages
- Backdoor Attacks on Multilingual Machine Translation
- BadLingual: A Novel Lingual-Backdoor Attack against Large Language Models
- Align Once, Benefit Multilingually: Enforcing Multilingual Consistency for LLM Safety Alignment
- Multilingual Safety Alignment Via Sparse Weight Editing
- Cross-Lingual Unlearning of Selective Knowledge in Multilingual Language Models
- P-MMEval: A Parallel Multilingual Multitask Benchmark for Consistent Evaluation of LLMs
- Multilingual Safety Alignment via Representation-Space Separability
- Backdoor Decontamination Dynamics in LLM Agents
- ICLShield: Exploring and Mitigating In-Context Learning Backdoor Attacks
- CL-Attack: Textual Backdoor Attacks via Cross-Lingual Triggers
- MELA: Multilingual Evaluation of Linguistic Acceptability