Metodología

Rocio Nefertari Garcia Luevano, Soto Contreras Luis Angel, Garcia Luevano Lorena Jacqueline, Lopez Fregoso Alexis, Ezbai Ahananel Guzman Rojas

Atentar contra la integridad de los niños, niñas y adolescentes por medio del reclutamiento forzado y el abuso sexual ha sido uno de los grandes problemas que se tiene en México, por ese motivo presentamos un modelo de IA que detecte conversaciónes de posible riego de reclutamiento forzado hacia al menor, utilizando una base de datos que contienen conversaciones clasificadas por nivel de riesgo que se dectecta dandole aviso a un padre de familia.

Reviewer's Comments

Reviewer's Comments

Arrow
Arrow
Arrow

Trabajos relacionados

La afirmación del paper de que no existen trabajos previos que aborden este problema es factualmente incorrecta y representa la debilidad más significativa de la entrega. Existe una literatura sustancial sobre detección de grooming infantil y comportamiento depredador en entornos de chat mediante NLP, incluyendo sistemas con arquitecturas casi idénticas, clasificación en tiempo real basada en transformers, alertas a padres, y entrenamiento sobre el dataset PAN12, que los propios autores citan. Ignorar esta literatura no fortalece la contribución; socava la credibilidad del paper. Los autores deberían reformular su sección de trabajos relacionados para posicionar su sistema dentro del estado del arte existente, destacando lo que es genuinamente novedoso, como el contexto latinoamericano regional, los diccionarios de argot criminal específicos de México y la integración específica con Roblox.

Afirmaciones de robustez

El paper hace afirmaciones fuertes sobre la robustez del sistema y su superioridad frente a enfoques alternativos, pero estas no están respaldadas por una evaluación empírica rigurosa. Los únicos resultados presentados consisten en una única conversación simulada con un puñado de mensajes, y notablemente, varios mensajes claramente peligrosos reciben puntuaciones de riesgo muy por debajo del umbral de alerta de 0.75, lo cual es en sí mismo un indicador preocupante de falsos negativos, no una evidencia de robustez. Los argumentos arquitectónicos sobre la tokenización WordPiece manejando ofuscación, o LoRA previniendo el sobreajuste, son teóricamente razonables pero insuficientes para sostener las afirmaciones realizadas sin estudios de ablación o evaluación sobre un conjunto de prueba de tamaño adecuado. Los autores deberían reemplazar estas aserciones con métricas concretas, precisión, recall y F1, calculadas sobre un dataset de al menos varios cientos de ejemplos, idealmente incluyendo casos adversariales con ofuscación deliberada de texto.

This is one of the most important problems in the batch, and it's worth saying why out loud. In Mexico, organized crime recruits kids directly, and the chat in an open game is now one of the doors they walk through: an adult builds trust fast, dangles money or Robux, and starts steering a child toward a meeting or a "job." It isn't only Mexico either. Grooming and recruitment over Roblox and Free Fire is a global pattern wherever kids play unsupervised, which is everywhere. A cheap parental-alert layer aimed at exactly that door could genuinely help a family, and you picked the right door. The architecture is the strongest part. The serverless alert plumbing (API Gateway, Lambda, S3, alerts out through SES and SNS) is concrete and sensible, it runs in the free tier, and the BPMN actually shows how a message travels from chat to alert. You know how to build the pipe. And the simulated Mario/Juan chat nails the grooming arc: fast intimacy, the Robux hook, then the probe for personal info. That's the exact sequence worth catching, and you saw it.

The biggest opportunity now is the evidence, because the paper describes a different system than the one in your repo. The write-up talks about a Transformer (DistilBERT) adapted with PEFT/LoRA and transfer learning, but the code that's actually shared is a keyword dictionary with a substring match, the simple kind of filter the paper positions itself above. That gap is the first thing to close, and it cuts both ways: either build the model you describe, or rewrite the paper around the keyword baseline you actually have and own it, because a transparent rule-based first pass is a legitimate starting point. The numbers have the same problem. Section 4C mentions a statistically significant recall gain over Naive Bayes with cross-validation, but the backing isn't there, no dataset size, no split, no precision/recall/F1, no table. And the thresholds don't agree: the paper says alerts fire at 0.75, the code says 0.05 with a clause that flags on a single keyword hit regardless of score, which is why messages scoring 0.028 still trip an alert. I'm flagging this plainly because for child safety it's the thing that matters most: a missed message is the entire cost, so the model, the scores, and the threshold all have to agree before anything else can be trusted. The good news is you're close to proving it. You already cite PAN12, which is a grooming-detection corpus, so run the model against it, report precision and recall and especially false negatives on a held-out split, and show the threshold actually firing on the messages you call risky. That one step turns "it works" into something a parent can rely on.

Cite this work

@misc {

title={

(HckPrj) Metodología

},

author={

Rocio Nefertari Garcia Luevano, Soto Contreras Luis Angel, Garcia Luevano Lorena Jacqueline, Lopez Fregoso Alexis, Ezbai Ahananel Guzman Rojas

},

date={

},

organization={Apart Research},

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

howpublished={https://apartresearch.com}

}

Recent Projects

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.

Read More

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.

Read More

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.

Read More

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.

Read More

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.

Read More

This work was done during one weekend by research workshop participants and does not represent the work of Apart Research.
This work was done during one weekend by research workshop participants and does not represent the work of Apart Research.