Accepted — IEEE Milcom 2026
Probing the Prefill: Detecting Code Vulnerabilities via Latent Activations
Khatri, A.
Companion writeup to our paper "Probing the Prefill: Detecting Code Vulnerabilities via Latent Activations" (arXiv:2608.16970)
Research
Peer-reviewed papers and preprints from the Wrynx team.
Accepted — IEEE DSN-W 2026
Latent Space Probing for Adult Content Detection in Video Generative Models
Khatri, A., Prabhu, C.
Video generation models can be coaxed into producing explicit or harmful content with little effort. We trained lightweight probes on the internal latents of CogVideoX to detect unsafe content before a single pixel is decoded — matching 8B guard models at over 1,000× lower latency.
Accepted — IEEE DSN-W 2026
Safety Beyond the Interface: Detecting Harm via Latent LLM States
Khatri, A., Prabhu, C., Neogi, O.
We kept running into the same frustration with external guardrails: they're blind. They see tokens — what went in, what came out — but nothing in between. So we asked a simpler question: does the model already know when a prompt is harmful? We trained lightweight probes on LLaMA-3.1-8B's hidden states and found that it does — matching 7B guard models at a fraction of the cost.
Preprint — arXiv
Same Probe, Different Numbers: Are Activation Probes Robust to Inference-Time Numerical Non-Determinism?
Khatri, A.
Activation probes are trained under one inference setup and deployed under another. Across three models and 768 probes, we find they barely notice batch-size and precision changes when reading the prompt — but during generation, the same rounding noise changes the text the model writes, and that is what flips probe verdicts.
Preprint — arXiv
Do All LLMs Know When They're Being Harmful? A Reproducibility Study of Latent-Space Safety Probes Across Model Families
Khatri, A., Chan, D. L.
A reproducibility study of latent-space safety probes across model families.