Browse categories
English
$
USD
Connect to your AI
I will test your ai chatbot or agent for prompt injection
Md. Azmol Haque
About this gig
Most AI security advice stops at "add a filter." That is not what stops an attack on an agent.
A chatbot produces text, so the worst case is bad output. An agent takes actions, so the worst case is a bad action. I report on what your system can be made to DO, not just what it says.
What I test:
- Direct prompt injection: instruction override, role confusion, system prompt extraction
- Indirect injection: hostile instructions hidden in documents or pages your AI reads
- Tool and function-call abuse: can input make it call a tool it should not?
- Data leakage: system prompt, keys, other users' context
- Guardrail bypass: where a filter is a tripwire, not a control
You get a written report: every confirmed finding with reproduction steps, severity and a specific fix. Written for engineers, not a scanner dump.
I build these systems too. My open-source agent pipeline defends injection architecturally - its extraction stage holds no network tools. My secret scanner ships 108 patterns across 805 tests.
Credited in Google's Vulnerability Reward Program and by UNICEF.
Message me before ordering and I will tell you honestly if it is in scope.
Get to know Md. Azmol Haque
Md. Azmol Haque
AI Safety Security Researcher
- FromBangladesh
- Member sinceJan 2023
- Avg. response time17 hours
Languages
Bengali, English
Security researcher and LLM systems engineer. Google VRP-credited for a disclosed authentication bypass, and acknowledged by UNICEF for a high-severity exposure affecting 13,000+ accounts.
I build my own tooling - an exposure scanner and an LLM security guardrail - so results are fast and reproducible. Every finding is verified before you see it: signal, not false positives.
I work only with written authorization.
English (fluent) - Bengali (native)
My Portfolio
FAQ
Will you break my chatbot or leave it in a bad state?
No. Testing runs on test accounts against non-production data, and I don't make changes to your system. If a successful attack triggers a tool call, I stop there and document it rather than repeating it.
What if nothing works - do I still get a report?
Yes, and that is a good outcome worth paying for. You get the full attack matrix: every family tested, how many prompts each, and what your defences caught. Knowing which attacks your system already resists tells you where not to spend next.
Do you need production access or customer data?
No. A test account or an API key against a staging deployment is enough. I never test against real customer data or a live assistant serving real users.
I only have a simple chatbot, not an agent. Is this worth it?
Yes, though the risk profile is different. Without connected tools the main exposures are system-prompt disclosure, policy bypass and reputational damage. Message me with what it does and I'll tell you honestly whether Basic covers you.
Can you test an AI product I don't own?
No. I need written authorization from whoever owns it. This is a rule I do not make exceptions to.
How does the free retest work?
I deliver the report first and the order completes on that. The retest is redeemable for 30 days afterwards - message me when your fixes are live and I will verify them and send an updated report. No new order needed.
