AI Security Researcher · Adversarial ML · Hardware Security · Compiler Security

Kathmandu, Nepal · MLCommons Contributor
manjitpokhrel.com.np · Email · LinkedIn · ResearchGate · Medium

I study adversarial safety failures in large language models and build systems-level tools for inference optimization and security analysis — from GPU kernels to compiler passes.

Currently investigating:

• Multilingual safety alignment failures in RLHF-trained LLMs
• Compiler-level safety degradation — does torch.compile silently break alignment?
• Cache-timing side-channel attacks on transformer inference
• Activation-level jailbreak detection via neuron firing patterns
• Training-free inference sparsity via custom CUDA/Triton kernels

Paper (Zenodo) · Code

First structured adversarial safety benchmark for Nepali-language LLMs.

• 1,200+ multilingual and code-switched adversarial probes
• 73.7% safety bypass rate in Nepali vs 0% in English
• Evaluated across Qwen, Gemma, Llama, and Gemini
• Findings disclosed to Google AI VRP (triaged) and Meta Whitehat
• Independently confirmed by Qwen engineer
• Revealed systemic tokenizer and alignment asymmetries in low-resource languages

PyPI · Code

Training-free activation sparsity framework for LLM inference.

• Custom CUDA kernels: sparse row-packing at warp/shared-memory level
• 110.5% inference speedup on RTX 5060 (Blackwell)
• Zero retraining. Zero fine-tuning. Zero accuracy loss.
• Currently porting to OpenAI Triton as GhostWeight-Triton (sm_120)

Adversarial attack vector I coined. Exploits morphological transformations in Devanagari script to bypass LLM safety filters at the subtoken level — below where safety alignment operates.

Investigating whether ML compiler optimizations (torch.compile, Triton kernels, TensorRT) silently alter safety alignment in RLHF-trained models. Zero prior work exists on this intersection.

Extending Flush+Reload side-channel attacks from CNNs (Cache Telepathy, USENIX 2020) to transformer attention mechanisms — exploiting input-dependent memory access patterns unique to self-attention.

• MLCommons — Contributor, DMLR & AI Risk and Reliability Working Groups (2026–present)

Security & Adversarial ML
Jailbreak evaluation · Cross-lingual safety bypass · Tokenizer exploitation · Morphological attacks · Side-channel analysis (cache-timing) · Responsible disclosure

ML Frameworks
PyTorch · HuggingFace Transformers · PEFT · NumPy · scikit-learn · llama.cpp

GPU & Compiler Systems
CUDA 12.8 · CUDA C (warp-level kernels) · OpenAI Triton (sm_120 Blackwell) · MLIR (learning) · Nsight Compute · Roofline analysis

Systems & Hardware
Flush+Reload (C) · Cache-timing attacks · AMD Zen 4 microarchitecture · · Operator fusion · Graph breaks

Languages
Python · C · CUDA C · JavaScript · SQL

Infrastructure
Linux · Git · Docker · FastAPI · HuggingFace Spaces · Vercel

RTX 5060 8GB (Blackwell, sm_120) · Ryzen 7 7700 (Zen 4) · 16GB DDR5

Kathmandu University
BSc Computer Science (Expected 2029)