Jun 22, 2026

Beyond English: Assessing the Robustness of LLM Safety Mechanisms Against Structural Jailbreaks in Spanish

Jonathan Narvaez

LLM safety guardrails are trained mainly to catch direct, single-turn harmful requests in English, leaving open whether structural attacks — long-context framing and multi-turn escalation — work just as well in Spanish, particularly in dialects like Bogotá Spanish. This matters for Latin America's emerging AI governance frameworks, which assume safety guarantees never verified for Spanish deployments. We adapted 16 AdvBench-style harmful instructions into long-context and multi-turn attack formats and tested them against Gemini's standard chat interface under a black-box threat model. An exploratory session combining fictional framing with multi-turn escalation produced harmful content across all three tested categories — identity theft, violence-inciting content, and discriminatory speech — and across a broader set of 16 sub-objectives, 62.5% showed some degree of safety bypass. Gradual escalation, not framing alone, drove compliance, with partial-compliance responses acting as a foothold for further escalation rather than a safety backstop. While this single-session pilot lacks statistical power and a control group, it offers an existence proof that Spanish-language structural jailbreak vulnerabilities are real, motivating controlled replication as a priority for regional AI safety evaluation.

Review Project

View Related Sprint

Reviewer's Comments

Length vs. Content

The paper is disproportionately long relative to the actual empirical contribution it delivers. A pilot study with a single session and no statistical results does not warrant the page count, and the redundant tables and repeated content suggest the length is largely artificial.

No real metrics/position paper

The paper presents itself as an empirical study but delivers no quantitative results in any defensible sense. The headline 62.5% bypass figure comes from one uncontrolled session with manual, non-blinded classification. It reads as a position paper and existence proof, which is a valid contribution, but should be framed honestly as such.

Missing control conditions

The methodology explicitly calls for comparing English, neutral Spanish, and Bogotá Spanish, but none of that comparison was actually conducted. The core research question of whether Spanish-language structural jailbreaks behave differently from English ones remains entirely unanswered.

Table formatting

The results tables have no numbers, no proper captions, appear duplicated across sections with minor variations, and are full of wasted whitespace. A single consolidated table would communicate the same information far more clearly.

What works

The limitations section is unusually honest and self-aware for a student paper, the related work is well-contextualized, and the research question is genuinely relevant to Latin American AI governance. The bones of a good future study are here.

The paper addresses an important and fascinating question regarding the success of structural jailbreaks in languages other than English, but the current experiments do not clearly establish whether jailbreak success actually differs across English, neutral Spanish, and more strongly marked dialectal varieties. The authors acknowledge this limitation themselves. However, without a direct comparison of attack success rates, it remains unclear whether the observed effects reflect changes in model vulnerability. While the paper is clearly presented and openly discusses its limitations, the experimental design does not directly test the central hypothesis. Given the limited duration of the hackathon, a smaller set of jailbreaks evaluated systematically across these language varieties might have provided stronger evidence for the paper's main claim and increased the practical relevance of the findings.

El proyecto aborda una pregunta relevante para la seguridad de modelos de lenguaje en contextos hispanohablantes: si los mecanismos de seguridad resisten ataques estructurales cuando se realizan en español, y no solo en inglés o frente a solicitudes dañinas directas. Su intuición principal me parece valiosa, no debería asumirse que un modelo es seguro para usuarios hispanohablantes si no ha sido probado en sus idiomas y formas reales de interacción, y su mayor mérito está en trasladar al español una familia de ataques ya conocida (contexto largo, ficción y escalamiento conversacional) y mostrar, en una interacción real, que esa vulnerabilidad puede aparecer. Sin embargo, el proyecto no propone una técnica nueva ni demuestra de forma controlada una vulnerabilidad específica del español bogotano frente al inglés o al español neutro: percibo que funciona mejor como una prueba exploratoria de existencia que como una evaluación comparativa sólida.

La principal debilidad está en la ejecución, y el propio equipo reconoce con honestidad varias limitaciones: es una sesión única, sin grupo control, que combina varias técnicas sin aislar cuál produjo el bypass, con clasificación hecha por el mismo equipo y sin probar mitigaciones. Valoro esa transparencia, pero las limitaciones no son menores. A esto se suma que el documento plantea una comparación entre inglés, español neutro y bogotano, pero los resultados visibles provienen de una sola sesión contra un solo modelo, que el propio texto diferencia de la comparación cuantitativa prometida, y que las cifras reportadas se contradicen entre sí, lo que impide reconstruir con claridad cuál es el resultado consolidado. La extensión del anexo no debe confundirse con una muestra amplia: documenta una interacción larga, no múltiples pruebas independientes.

Desde mi eje de auditoría y rendición de cuentas, percibo un punto especialmente importante, porque toca la coherencia entre lo que el trabajo predica y lo que practica. El documento afirma, al describir su sesión exploratoria, que no reproduce las salidas; sin embargo, el anexo incluye salidas transcritas como "verbatim", extensas y asociadas a categorías sensibles. Aun si correspondieran a tandas distintas, el documento no establece esa distinción ni justifica por qué unas se manejan con contención y otras se publican íntegras. Para un trabajo centrado en seguridad, sentí que esta tensión debería corregirse con una política clara de divulgación responsable. En esa misma línea, el anclaje regulatorio se queda en la mención: invoca marcos regionales para justificar su relevancia, pero no identifica qué principio o responsabilidad institucional concreta queda tensionada por el hallazgo.

Para fortalecer el proyecto, el siguiente paso sería convertir esta alerta en un protocolo más controlado y reutilizable: comparar los mismos ataques en inglés, español neutro y variantes locales; ampliar modelos y categorías de daño; separar el efecto de cada técnica; incorporar jueces independientes; probar mitigaciones mínimas; y establecer una política de divulgación responsable que el propio reporte cumpla.

Y quiero dejar una reflexión, con respeto por el trabajo real que hay detrás y por la honestidad con que el equipo reconoce sus límites: la tesis del proyecto (que la seguridad no puede afirmarse sin verificarla), me parece valiosa, y justamente por eso percibo que es una invitación que el reporte podría hacerse también a sí mismo, aplicando a su propia evidencia el mismo rigor que reclama para los modelos. En conjunto, el proyecto tiene valor como primer paso y como advertencia frente a la transferencia automática de garantías de seguridad del inglés al español; su aporte debe mantenerse proporcional a la evidencia, que todavía es exploratoria.

Cite this work

@misc {

title={

(HckPrj) Beyond English: Assessing the Robustness of LLM Safety Mechanisms Against Structural Jailbreaks in Spanish

author={

Jonathan Narvaez

date={

6/22/26

organization={Apart Research},

note={Research submission to the research sprint hosted by Apart.},

howpublished={https://apartresearch.com}

}

Recent Projects

View All

Apr 27, 2026

OliGraph: graph-based screening of large oligopools

Existing synthesis screening tools cannot evaluate short oligonucleotide pools, whose overlapping fragments can be reassembled into regulated sequences via polymerase cycling assembly (PCA) yet fall below gene-length detection thresholds. We present OliGraph, an open-source tool that constructs a bi-directed overlap graph from an oligonucleotide pool and extracts contigs for downstream gene-length screening. An optional PCA mode retains only cross-strand overlaps consistent with PCA chemistry. We validated OliGraph in a blinded study across ten simulated pools (70–9,184 oligonucleotides, 30–300 bp) spanning four risk categories. BLAST screening of individual oligonucleotides failed to identify sequences of concern in most pools: three returned zero hits, and vector noise obscured true positives in the remainder. After OliGraph assembly, contig-level BLAST matched the longest assembled sequences (up to 1,905 bp) to sequences of concern at 97–100% identity. In one pool, assembly collapsed 1,634 individual BLAST results into 10 hits from a single contig, all assigned to the same source organism. PCA mode correctly distinguished assemblable from non-assemblable fragments within the same pool. Two pools with no assemblable structure yielded no contigs. OliGraph processed all pools in under 0.2 seconds, fast enough for real-time order screening and consistent with proposals to bring oligonucleotide orders within the scope of synthesis screening regulation.

Apr 27, 2026

BioRT-Bench: A Multi-Attack Red-Teaming Benchmark for Bio-Misuse Safeguards in Frontier LLMs

Frontier AI laboratories are expected to maintain safeguards against biological misuse, but whether deployed models actually refuse bio-misuse queries under adversarial pressure is largely unmeasured in the public literature. We introduce BioRT-Bench, a benchmark that runs four attack methods (direct request, PAIR, Crescendo, and base64 encoding) against four frontier models (Claude Sonnet 4.6, GPT-5.4, DeepSeek V4-flash, Kimi K2.5) across 40 prompts spanning five biosecurity-relevant categories. Responses are scored by a calibrated judge extending StrongREJECT with two bio-specific dimensions: specificity and actionability. We measure Attack Success Rate (ASR), where 0 means the model fully refused and 1 means it provided specific, actionable bio-misuse content. Our results reveal a sharp robustness divide: Chinese frontier models (DeepSeek, Kimi) have under 5% refusal rates even under direct request (ASR 0.88 and 0.79), while Western models (Claude, GPT) maintain substantially stronger safeguards (ASR 0.15 and 0.16). Crescendo is the most effective attack across all models, both in bypassing refusal and in eliciting actionable content. Claude Sonnet 4.6 is the most robust model tested, achieving 100% refusal against base64-encoded prompts.

Apr 27, 2026

PROTEUS (PROTein Evaluation for Unusual Sequences): Structure-Informed Safety Screening for de novo and Evasion-Prone Protein-Coding Sequences

AI protein design tools like RFdiffusion, ProteinMPNN, and Bindcraft make it trivial to produce low-homology sequences that fold into active, potentially hazardous architectures. However, sequence homology-based biosafety screening tools cannot detect proteins that pose functional risk through structurally novel mechanisms with no sequence precedent. We present a tiered computational pipeline that addresses this gap by combining MMseqs2 sequence alignment with structure-based comparison via FoldSeek and DALI against curated toxin databases totaling ~34,000 entries. AlphaFold2-predicted structures are screened for both global fold similarity (FoldSeek) and local active/allosteric site geometry (DALI), capturing convergent functional hazards that sequence screening misses. The pipeline was validated against a panel of toxins, benign proteins, structural mimics, and de novo-designed Munc13 binders, as well as modified ricin variants with residue substitutions. We additionally tested robustness to partial-synthesis evasion, where a bad actor submits multiple shorter coding sequences intended for downstream reassembly into a full toxin-coding gene. We found that while sequence-based screening did not identify any de novo ricin analogues with high certainty, the combined pipeline with FoldSeek and DALI identified all 24 tested de novo ricins as toxic.

Apr 27, 2026