Vishnu Kosuri

AI security researcher and self-taught AI/ML engineer.

I was the founding ML engineer at Vahini Technologies, where I built the machine learning behind a patented IMU handwriting pen. That work led to 2 granted Indian patents, and I am the lead inventor on both.

I work hands-on with models: training and fine-tuning them, reinforcement learning, and evaluations. My current research is on jailbreaks, LLM red teaming and prompt injection.

Open to AI safety and AI security research roles, and to offensive security. Open to remote and international opportunities.

Work

  • Aug 2026 to now

    Founder, Zenith

    I built Zenith, an AI-powered security assessment platform with more than five early clients, alongside a vulnerability scanner using a local open-weight model and workflows for testing coding agents.

  • Dec 2024 to Aug 2026

    Founding ML Engineer, Vahini Technologies

    Employee #1. Owned the AI, ML, data and software stack for the handwriting pen and built its machine learning pipeline, work that led to 2 granted patents. Shipped and secured the web app, REST APIs and PostgreSQL backend. Reverse engineered the pen's firmware with Ghidra and found hardcoded credentials in two shipping variants, then worked with the team on the fix.

  • 2024 to now

    Research Analyst, CSAI Research and Development Cell

    Industry-advisory research on AI security under Lt. Gen. (Dr) Rajesh Pant, India's former National Cyber Security Coordinator. I write explainers on adversarial threats to AI systems and help run events that have reached more than 600 professionals.

  • Jul to Oct 2024

    Security Consultant, Tapnex

    Found 11 vulnerabilities in a cashless-event platform's web, mobile and API surfaces, including an MFA bypass and IDOR-based privilege escalation, and retested the fixes before launch.

  • Jun to Jul 2024

    Penetration Testing Intern, ISAC National Security Database

    Found 18 vulnerabilities, 5 of them critical, across two commercial production sites during authorized assessments.

Security research

My security work spans web, mobile and API penetration testing, vulnerability assessment, secure code review and product security. I'm also interested in firmware reverse engineering and the security of connected devices.

  • Orca Whirlpools, Immunefi #65797: reported an out-of-bounds account-buffer access and received a USD 500 bounty.
  • Folks Finance, Immunefi #69077: submitted a staking-cap accounting finding that could lock deposits.
  • Gray Swan: 47th in the June 2026 indirect prompt injection challenge.

CTF results

  • 1stSANS NetWars Tournament Core
  • 1stOWASP AppSec Bangalore BeSec CTF
  • 4thCloud Security Alliance Bangalore CTF
  • 6thSANS Veterans Day CTF
  • 7thAmazon CTF

Credentials

  • GIAC GSECSecurity Essentials
  • GIAC GCIHCertified Incident Handler
  • CEH v12Certified Ethical Hacker, EC-Council
  • eJPTJunior Penetration Tester, INE Security
  • GIAC GFACTFoundational Cybersecurity Technologies
  • ISC2 CCCertified in Cybersecurity

Scholarships and grants

  • SANS Institute scholar, SANS Technology Institute
  • USD 25,000 grant, Y Combinator Startup School India, 2026

Programs

  • Compute Governance Career Accelerator, Cohort 1, 2026
  • 1752 Ventures Accelerator, 2026
  • Founder Inc. Canopy, 2025

Community

  • Founded Salus, Jain University's cybersecurity club, now 200+ members
  • AWS Cloud Captain: 4 cloud security workshops for 150+ students
  • Cyber Security Facilitator, GDG On Campus

The pen and its patents

As the founding ML engineer and employee #1 at Vahini Technologies, I owned the AI, ML, data and software stack for an IMU-based handwriting pen. You write on ordinary paper, and the pen works out what you wrote from how it moved.

Patent certificate No. 584433 from the Patent Office, Government of India
Patent No. 584433Handwriting recognition with an intelligent ballpoint pen using IMU sensors. Granted 23 March 2026.
The IMU handwriting pen, with its circuit board and ink cartridge shown above the assembled pen
The pen, with its sensor board taken out.

I built the machine learning pipeline end to end: sensor-data pipelines, noise reduction, and CNN and Bi-LSTM recognition models, then full-stack development and product integration. We ran pilots in 5 schools with more than 500 students. I also worked with Andhra Pradesh government stakeholders and served as liaison for our IIT Tirupati collaboration.

This work contributed to 2 granted Indian patents, covering handwriting recognition and structured digitization. I am the lead inventor on both.

Patent certificate No. 604895 from the Patent Office, Government of India
Patent No. 604895A hybrid handwriting system for real-time digitization and structured data processing. Granted 1 October 2026.

1Pen on paper

2Raw motion signal

What the sensor reports: three accelerometer axes, with noise and slow drift.

3Noise reduction

Drift removed, then a Kalman filter on each axis. The box is the window the model is reading.

4Features

Stroke direction in 8 bins, plus speed and pen contact, for each short window of time. This is what the CNN sees.

5CNN and Bi-LSTM decoder

Computed live in your browser from the strokes of my name. The sensor noise is simulated and the decoding is staged for illustration; the real model was trained on recorded pen data.

Research interests

  • Jailbreaks
  • LLM red teaming
  • Prompt and indirect prompt injection
  • Agent tool-use security
  • Agent sandboxing
  • Adversarial robustness
  • Adaptive safety evaluations
  • Benchmark validity
  • Securing ML infrastructure
  • Fine-tuning
  • Reinforcement learning
  • Interpretability

These are my current interests, and I don't update this list often. If any of it overlaps with your work, I'm always open to a conversation or a coffee at the intersection of AI, cyber and hardware.

Say hello

Pick a 15-minute slot on my calendar, or email me. I read every email that isn't a sales pitch.

kvr.vishnu23@gmail.com