Thamilvendhan Munirathinam

Independent Researcher

Email: mthamil107@gmail.com


About

I work on the governance, safety, and security of LLM agents — in particular, lightweight in-band mechanisms that let infrastructure communicate policy to autonomous agents, and empirical measurement of whether agents honor such signals.

Writing

2026 · In-Band Signal Compliance (IBSC)
Two failures — obeying an illegitimate in-band signal, and ignoring a legitimate one — are mirror images on a single legitimacy axis. A benchmark (signalbench) measures both across six frontier models; none exceed 0.85.

Research interests

Publications

Will the Agent Recuse Itself? Measuring LLM-Agent Compliance with In-Band Access-Deny Signals
Preprint, 2026 · arXiv:2606.06460
A published mini-standard — the Recuse Signal — that lets a server ask a connecting LLM agent to voluntarily withdraw, with the first measurement of whether compliant agents honor it. Code: github.com/mthamil107/Recuse.
memorywire: A Vendor-Neutral Wire Format for Agent Memory Operations
Preprint, 2026 · arXiv:2606.01138
A vendor-neutral wire format for LLM-agent memory operations, for interoperability across agent frameworks and runtimes.
Beyond Pattern Matching: Seven Cross-Domain Techniques for Prompt Injection Detection
Preprint, 2026 · arXiv:2604.18248
Cross-domain techniques for detecting prompt-injection attacks against LLM-integrated applications.

Also on Google Scholar.

Background

Alongside research, I build and operate AI/agent systems in production:

Links