Jun 21, 2026

Hey Muslims, these LLMs may think you are a terrorist

Isra

I explored geopolitical origins of anti-Muslim bias, specifically how country of model origin shapes islamophobic stereotypes across French, American, and Chinese LLMs, and across two languages (French and English). The prompts related to three countries with controversies regarding the treatment of their muslim minority populations, specifically France, China, and India.

Review Project

See Code

View Related Sprint

Reviewer's Comments

The question is interesting and feels new to me. Looking at the country where the model comes from, together with the language and the context, is a smart idea. I checked your code and data and the pipeline is real, you did a lot of actual work here. I also want to thank you for being honest. You did this alone, with bad wifi, and you explain every limit very clearly. The main problem is that the results are more like impressions than proof. There are no statistical tests, so saying that one model is the most biased is not safe yet. The test questions were made by an AI and were not really checked, so when a model passes it can simply mean the question was weak. And one AI is judging a study that is about bias, so please check that judge against your own reading on a few examples. Good direction, it just needs stronger evidence.

The project had real data and results, but the conclusions still felt early-stage. To improve it, I would make the testing process more consistent, rerun any failed cases, and manually check some of the judging results. This would make the findings easier to trust and would help separate interesting patterns from possible noise in the data.

Good work. Comparing models built in different countries to see whether origin shapes their behaviour is a fresh angle that most bias studies skip, and you backed it with a real pipeline across three models and two languages. Your strongest insight is that bias takes more than one form. A model can cause harm by producing a stereotyped response or by refusing to engage at all, and a blanket refusal can look safe while really being avoidance. That is a sharp observation.

Where it needs strengthening is the evidence. The automated judge failed on about a quarter of the runs, so the findings rest on fewer cases than the totals suggest, and without significance testing it is hard to know which differences are real. One claim, that the Chinese model suppresses Uyghur related prompts, is not supported by your own data, where Muslim prompts were refused more often and the sample was very small. I would prioritise recovering the failed judge runs, adding basic significance testing, and revising the claims the data does not back. Using a single model as the judge also lets its own blind spots shape the results, so a few human checks would help.

Cite this work

@misc {

title={

(HckPrj) Hey Muslims, these LLMs may think you are a terrorist

author={

Isra

date={

6/21/26

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