Research

Published work on AI safety, adversarial robustness, and machine learning systems.

Papers

Perturbation-Based Generation Profiling Detects Covert AI Agent Attacks Where Token-Level Statistics Fail

Yatharth Maheshwari, Arka Dash, Abhineet Som

Unsupervised attack detection via cross-perturbation generation profiling — 0.94–1.00 AUROC across six LLMs (8B–32B) with zero training data, where token-level baselines plateau at 0.56–0.83.

HGT Leaves a Linear Fingerprint in Codon Space

Arka Dash, Yatharth Maheswari

Nucleotide virulence factor benchmarks are inflated by ~0.30 AUROC from organism confounds and gene-family leakage.

BioChain: Cross-Vendor Threat Detection via Function-Aware DNA Fragment Screening

Arka Dash, Yatharth Maheshwari, Asutosh Rath , Caio Timm, Igor Pereverzev

Codon frequency + logistic regression matches deep models genus-to-genus once you fix the data leakage that inflated every prior benchmark by ~0.30 AUROC.

Recent Writing

Nothing here yet — check back later.