Home

Ansar Shaikh

AI security engineer: testing what LLM apps do with untrusted input

I attack small LLM and RAG apps with prompt injection, build guards around them, and measure whether the guards help.

Projects

  • Injection success 3/8 to 1/8

    rag-redteam

    Indirect prompt-injection tests against a local RAG app on qwen2.5:1.5b with three defences. Spotlighting and a regex sanitiser changed nothing alone; the output guard did the work.

    • Python
    • Ollama
    • OWASP LLM01
  • 19 findings, 0 on clean servers

    mcp-audit-lite

    Static scanner for MCP server manifests and source: tool poisoning, hardcoded secrets, shell=True and risky permissions. It flags all 5 vulnerable samples and stays quiet on the 2 clean ones.

    • Python
    • pytest
  • Hindi and Marathi injections

    indicguard

    Prompt-injection detection for Hindi, Hinglish and Marathi input. An English-only regex caught none of the Hindi or Marathi attacks; IndicGuard reached F1 0.824 on a held-out set.

    • Python
  • Found and fixed a card regex bug

    llm-redactor

    FastAPI middleware that redacts PII and secrets, including PAN, UPI and IFSC, from LLM output. The evaluation caught a card-number regex bug, which I fixed.

    • Python
    • FastAPI
  • Local-LLM alert triage

    l1-copilot

    Sends Wazuh SSH alerts to a local LLM, with input sanitising against prompt injection. On 100 synthetic alerts the 1.5B model only reached recall 0.105, and the README says so.

    • Python
    • Ollama
    • Wazuh

More on GitHub

runbook-bot · secrets-leak-guard

Skills

  • LLM security: Indirect prompt injection, OWASP Top 10 for LLM apps, spotlighting, output guards, PII redaction, MCP server review
  • Evaluation: Attack-success rate, held-out test sets, precision and recall, local models with Ollama
  • Engineering: Python, FastAPI, pytest

2025 B.E. Information Technology, Sinhgad Institute of Technology & Science (SPPU), Pune, with Honours in Cybersecurity.

Experience

  • Aug 2025 – Present

    CTF Player and Security Researcher · Hack The Box and TryHackMe

    Hack The Box and TryHackMe labs on Linux and Windows. Thinking like the attacker is the part that carries over to red-teaming LLM apps. Placed 13th in the CloudSEK CTF (Aug 2025).

    • Offensive testing
    • Reporting
  • May – Jun 2024

    ServiceNow System Administrator, Virtual Intern · AICTE program

    Built ServiceNow ITSM reports and dashboards and completed the CSA labs.

    • ServiceNow

Certifications

  • Oct 2025

    Certified Ethical Hacker (CEH) Practical

    EC-Council · Credential ID ECC9356741028

  • Mar 2024

    Certified AppSec Practitioner (CAP)

    The SecOps Group · Credential ID 8481909

Other roles: SOC analyst · Cloud security & DevSecOps · DevOps / SRE / NOC · VAPT & AppSec · Software & QA · Security (SOC, VAPT and AI security) · Writing