Karthikeyan Sadayamuthu

Karthikeyan Sadayamuthu

Autonomous Systems Security

ORCIDGitHubLinkedIn

10Open frameworks
9Zenodo DOIs

Karthikeyan Sadayamuthu

I study how autonomous AI agents fail, and build open tools that check their work before it ships: who they act for, what they change, and what they leak.

Through OpenAstra, I publish open, citable frameworks that make an agent's permissions and actions verifiable before they reach production: policy languages, compliance-as-code tools and pre-merge checks. My paper on hunting attacks against LLM agents with SQL over cross-layer telemetry is accepted at IEEE CSTA 2026.

I bring sixteen years of building secure, large-scale software to this work. At Autodesk, I am a Principal Software Engineer and lead architect of the Trust platforms across Commercial, FedRAMP Moderate and AWS GovCloud, including a multi-agent platform built on MCP that enforces authorization at the data source, not in the model.

I review for Cloud Security Alliance working groups on AI safety, agent runtime management, identity and Zero Trust.

Research

OpenGPL

A declarative policy language for generative AI agents: what they can access, what they can produce, and the evidence to prove it.

anySQL

Query LLM responses, agent traces and RAG pipelines with standard SQL.

ControlGate

Checks code changes against NIST 800-53 at pre-commit and PR time, with block or warn verdicts and remediation guidance.

ExposureGate

Blocks a change from adding a dependency version with a known, exploited vulnerability, by joining OEBC to the OSV database.

OCBCOpen Controls Baseline Catalog

NIST SP 800-53 Rev. 5 and FedRAMP controls as a machine-readable catalog you can build on.

OEBCOpen Exposure Baseline Catalog

A CVE exposure baseline enriched with EPSS scores and CISA KEV data.

unasked

Scores each change a coding agent made against your prompts, then reverts just the ones you never asked for.

mergelock

Catches conflicts git can't see between parallel AI agent branches, such as changed signatures with new callers.

ZeroEgress

An offline AI code editor using local models, so nothing leaves your machine.

OpenAPL

Access policy language for humans and AI agents.

Browse by line of work →

Publications

Untraced by Design: Hunting Attacks on LLM Agents with SQL over Cross-Layer Telemetry

IEEE International Conference on Cybersecurity Systems, Technologies and Applications (CSTA 2026)

2026

All publications →

Peer review & judging

Designing the AI-First Security Organization

Cloud Security Alliance · AI Safety WG · Peer reviewer · Nov 2026

Published

Navigating Identity and Access Management (IAM): Standards and Protocols

Cloud Security Alliance · Identity and Access Management WG · Peer reviewer · July 2026

Published

Build a Zero Trust Architecture

Cloud Security Alliance · Peer reviewer · Nov 2026

In review

The Six Pillars of MCP Security

Cloud Security Alliance · Peer reviewer · Nov 2026

In review

AARM Benchmarking Framework

Cloud Security Alliance · Peer reviewer · Oct 2026

In review

Authorization of AI Agents

Cloud Security Alliance · Peer reviewer · Oct 2026

In review

Securing Operational Technology in the Cloud with Zero Trust

Cloud Security Alliance · Peer reviewer · Oct 2026

In review