Bengaluru, India
My work spans software engineering, product security, machine learning and AI security research.
I was Vahini's first engineer, helping turn a sensor pen into a working handwriting-recognition product. Now I’m building Zenith and studying how AI systems fail when they read untrusted content or use tools.
I’m building ARES RED TEAM to connect security testing with the evidence behind each finding. The work includes coordinating tools, reviewing agent actions and making results easier to reproduce.
The Pi session evaluator checks a recorded agent session against a tool policy and points each finding back to the relevant action. The report is repeatable without running those actions again.
I worked across the handwriting model, web application, backend and product security. That meant collecting and preparing sensor data, training recognition models, building REST APIs and PostgreSQL services, and handling authentication, access controls and deployment.
I also examined device firmware with Ghidra, identified hardcoded credentials across two variants and worked with the team on remediation. The product reached pilots across five schools and more than 500 students.
Alongside engineering, I supported product demonstrations, investor conversations and the practical decisions needed to move beyond a prototype.
I reviewed a cashless-event platform across its web, mobile and API components. I found 11 issues, including authentication and access-control failures, documented the impact and worked with engineers to check fixes before launch.
I contributed to assessments of two commercial websites that identified 18 vulnerabilities, including five rated critical. My work included reproduction steps, severity assessments and remediation guidance.
Indian Patent 584433 · Granted 23 March 2026
I’m the first-named inventor on Handwriting Recognition with an Intelligent Ballpoint Pen Using IMU Sensors, developed with my team at Vahini Technologies.
The pen records motion through inertial sensors. I worked on data preparation, Kalman filtering, sequence alignment and CNN/Bi-LSTM recognition, as well as the software around the model. During development, the model reached approximately 82% character accuracy.
I’m interested in what an AI system actually did, not just how convincingly it describes its behavior. Did it expose information, take an unwanted action or prevent someone from completing an ordinary task?
My current work includes TAINT, which compares prompt-injection detectors; REFUSAL-CLIMB, which studies refusal signals; and AUTOFORGE, which runs budgeted disclosure experiments. I also built an AI vulnerability scanner that brings code-analysis findings and optional model explanations into a review workflow.
One recent correction was simple but important: a model that says “I won’t reveal the secret” and then prints it has still leaked it. I changed AUTOFORGE’s grader to score the disclosure separately from refusal wording.
As a research analyst, I contribute work on AI security and emerging threats for industry and public-service audiences. I have also supported events reaching more than 600 professionals.
Appointment announcementThrough Immunefi, I reported an account-buffer bounds issue in Orca Whirlpools and received a $500 bounty. I also submitted a finding concerning staking-cap accounting and deposit lockout in Folks Finance.
Immunefi researcher: VK05UR1. I placed 47th in Gray Swan’s June 2026 Indirect Prompt Injection challenge.
My contributions include file handling, resource cleanup, text processing and regression coverage.
I’m studying Computer Science and Engineering at Jain University, graduating in May 2027. I founded Salus Cybersecurity Club, which grew to more than 200 members. As an AWS Cloud Captain, I delivered four workshops for more than 150 students, and served as a Cyber Security Facilitator with GDG On Campus.
Campus facilitator announcementGIAC GCIH, GSEC and GFACT; CEHv13; eJPTv2; ISC2 CC. ISO/IEC 27001 Lead Implementer training completed.
Say hello
I’d love to hear about AI security research, early engineering roles and useful things we could build together.
kvr.vishnu23@gmail.com
