Portrait of Yuvraj Pradhan
Download CV Get in touch

Yuvraj Pradhan

AI Security Researcher and Engineer

I break AI agents at the trust boundaries they were never built to defend, and I build the security controls that close them.

Download CV Get in touch
LinkedIn ↗ Google Scholar ↗ GitHub ↗ Vyrion ↗
About

AI agents now act with real authority. I make sure attackers cannot borrow it.

I work on AI agent security. I find where agents trust state, tools and instructions they should not, and I build controls that hold under attack. My work defined Approval-Provenance Failure, an attack class where an agent accepts a rewritten database value as proof of human approval. I disclosed it to 29 vendors including Microsoft, OpenAI, Google, AWS and Anthropic, reported it to CERT/CC and the AI Incident Database (AIID), and built Vyrion, the open-source guard that blocks it. My current work, Instruction-Authority Divergence, shows coding agents obeying instructions that no reviewer approved.

First authorACM AISec 2026, with ACM CCS
ReporterCERT/CC case VU#550098
SpeakerBlack Hat MEA 2026, Riyadh
FounderVyrion, approval integrity
43systems evaluated, version pinned
29vendors and projects in coordinated disclosure
9talks across 5 countries
17+merged PRs to PyTorch, NumPy, CPython and more
Research

Approval-Provenance Failure ↗

2026 First author

An organisation can require an authorised human decision while the runtime accepts mutable workflow state as proof that the decision happened. Approvals can then be forged, rebound, replayed, suppressed or laundered before anything executes.

29 of 43systems exposed
29vendors disclosed
CWE-345Flowise advisory, Medium
CERT/CC VINCEMicrosoft MSRCGoogle AI VRPAWS VDPAnthropicASF SecurityAI Incident Database

Instruction-Authority Divergence ↗

2026 First author

The instructions a reviewer approves are not always the instructions an AI coding agent obeys. Dependency-supplied instruction files can override or travel past what a human actually reviewed.

200/200precedence runs won by a dependency
2,470repositories measured
30.5%file-level review gap
Publications
  1. Accepted

    Ghost Approvals: Approval-Provenance Failure in Agentic AI Frameworks

    Y. Pradhan, A. Singh, A. Kumari, A. Pradhan, D. Singh. 19th ACM Workshop on Artificial Intelligence and Security, co-located with ACM CCS 2026.

  2. Published

    Nano Large Language Model of 125 Million-Parameter for STEM Education

    Y. Pradhan, A. Singh, N. Kulkarni, A. Priyadarshi, O. Rajale, K. Mohite. Discover Computing, vol. 29, article 161, 2026. doi.org/10.1007/s10791-026-10033-z

  3. Under review

    Nobody Approved This: Approval-Provenance Failure Across Agentic AI, Workflows, and CI/CD

  4. Under review

    Who Witnesses the Human in the Loop? The Approval Record the System Can Write for Itself

  5. Under review

    The Agent Obeys the Wrong Principal: Instruction-Authority Divergence in Coding Agents

Talks
Oct 2026AltSecCon Bengaluru
  • Nobody Approved This: Hijacking Approval Authority
AltSecCon speakers and attendees in Bengaluru
Yuvraj presenting approval-provenance attacks at AltSecCon
Yuvraj on stage at AltSecCon asking why human approval matters
Jun 2026MCP Dev Summit Mumbai
  • The Sticky Session Bottleneck: Why MCP Agent Orchestration Breaks When You Scale
  • Who Do You Trust? Securing Multi-Agent MCP Systems in Healthcare
Audience at MCP Dev Summit Mumbai
Yuvraj on stage at MCP Dev Summit Mumbai
Yuvraj speaking at MCP Dev Summit Mumbai
  • 2026
    Lightweight Small Language Models for Low-Resource DevicesICRTEBTM 2026
The Stdio Deadlock Nobody Warned Us AboutWatch on YouTube ↗
Who Do You Trust? Securing Multi-Agent MCP Systems in HealthcareWatch on YouTube ↗
The Sticky Session Bottleneck: Why MCP Agent Orchestration Breaks When You ScaleWatch on YouTube ↗
Software

Vyrion ↗

Founder Apache 2.0

A fail-closed guard that enforces human-approval policy with an Ed25519-signed, single-use Seal bound to the approver, action, arguments, run, policy, expiry and nonce, instead of trusting mutable workflow state.

PythonNode.jsEd25519RFC 878515 frameworksPyPI

Ovrin ↗

Apache 2.0

An offline governance tool for coding-agent instruction files. It traces where each instruction came from, who reviewed it and which agents can reach it, then emits SPDX-aligned manifests and CI gates.

PythonSPDX 3.0GitHub ActionsCODEOWNERS
Open source

Merged documentation contributions to the libraries the AI field runs on.

5Hugging Face Transformers
4NumPy
2PyTorch
6+CPython, pandas, LangChain, scikit-learn
17+merged pull requests in total
2tools released, Vyrion and Ovrin
Education and expertise
MIT ADT University
B.Tech., Computer Science and Engineering, AI and Edge Computing
CGPA 8.26 Expected July 2027
AI security
Human oversight, authorization provenance, instruction authority, coordinated disclosure, threat modeling, trust-boundary analysis
Assurance
Ed25519, RFC 8785, SHA-256 commitments, replay defense, SBOM, SPDX 3.0, in-toto, CWE, CVSS
Agentic systems
LangGraph, CrewAI, OpenAI Agents SDK, Google ADK, AutoGen, Haystack, LlamaIndex, Semantic Kernel, Apache Airflow, MCP
Engineering
Python, JavaScript, Node.js, C, C++, SQL, PyTorch, Hugging Face, Docker, AWS, Google Cloud, GitHub Actions

Let's build AI that still listens

Open to research, engineering roles, collaborations and speaking. I reply within two working days.

Message on LinkedIn ↗ Download CV