Utsav Agarwal

Computer science at the University of Michigan

Experience

  • May — Aug 2026

    Faberlens AI

    AI security research intern

    • Replicated AI agent vulnerabilities previously confined to sandboxes against production-grade MCP servers, including Notion and Google Workspace, demonstrating real-world exploitability across live agentic pipelines.
    • Ran adversarial behavioral evaluations of agent skills, mapping the security regressions that appear when skills are composed.
  • Jan — July 2025

    Audium Health

    Software engineer intern

    • Analyzed data from 1,500+ patients by generating spectrograms, using ROC curves and visual metrics to improve model performance.
    • Automated management of 50k+ audio files to streamline supervised learning workflows, and built a Python visualizer deployed on Azure that gave customers real-time modeling results.

Projects

  • Oct 2026

    No Badge Actors

    A hardware security key for payments, built on DEF CON badges. It checks who you're paying, that they're standing in front of you, and what you're actually signing, before a button press releases the signature.

    Won Best Use of Solana · Won the Judged by an LLM track · MHacks 2026

    Solana, Ed25519, ESP32 / ESP-NOW, Node.js, React, TimescaleDB, Capital One Nessie

    Read the write-up →
  • Metadata Mirage

    Chained a web-app SSRF into AWS instance-credential theft and a single-API-call privilege escalation to administrator, then severed the whole chain with one instance-metadata flag — enforcing IMDSv2. Reproducible end to end with Terraform.

    SSRF to AWS administrator · Killed by one IMDSv2 flag

    AWS, IAM, Terraform, cloud security

    Read the write-up →
  • Sept 2025 — Apr 2026

    MRacing dashboard

    Reverse-engineered a Python driver dashboard and rebuilt it in C++, pulling live telemetry off the vehicle CAN bus for a Formula SAE team.

    ~3× faster · 40% less memory · 60+ FPS on a Raspberry Pi

    C++, Raspberry Pi, CAN bus

  • dashcam-perception

    Fine-tuned YOLOv8 on BDD100K so a dashcam model could detect the classes COCO never covers, and built the evaluation framework to prove it worked.

    +49.7% mAP@0.5 · riders and traffic signs learned from zero · 10 classes

    Python, PyTorch, YOLOv8

    Read the write-up →
See all projects →

Education

May 2028

University of Michigan

B.S.E. Computer Science · GPA 3.93 / 4.00

Relevant coursework: data structures & algorithms, computer organization, autonomous electronic systems, applied machine learning, mathematics of blockchains and cryptocurrencies.

LanguagesC++, C, Python, Java, JavaScript, SQL
Machine learningPyTorch, TensorFlow, OpenCV, NumPy, Pandas
EmbeddedArduino, Raspberry Pi, ESP32
CertificationsGoogle Cybersecurity Certificate

Contact