Creative Prompt Design for LLMs

Job ID: 40087014

Budget: ₹12,500 – ₹37,500 INR

Background

We are building a high-security online MCQ assessment platform intended for academic institutions and recruitment use cases.

A growing cheating pattern has emerged that traditional proctoring, browser lockdowns, and monitoring cannot fully prevent:

Candidates use a secondary mobile phone.They take a photo of the question on screen. Upload the image to AI tools (ChatGPT, Gemini, Claude, etc.). Receive the answer instantly and submit it

This bypasses:
- Browser restrictions
- Screen recording detection
- Webcam proctoring
- Network monitoring

We want to detect or deter this behavior at the AI-interaction level.

Objective

Design a reliable technical mechanism that allows us to detect when an assessment question image is uploaded to an LLM, by forcing or strongly encouraging the LLM to interact with a controlled external resource (URL / endpoint / QR / embedded signal).

The core detection idea:

If an LLM processes the question image, it should attempt to visit or resolve a URL (or equivalent signal) embedded within the question.
When this happens, our server receives a hit → we flag the candidate as a cheating suspect.

Key Challenge

- Most major LLMs do NOT fetch URLs embedded in images or text.
- QR codes inside images are often ignored.
- Even when URLs are visible, models typically do not perform outbound requests.

What We Need

- We want to break or bypass this limitation, legally and ethically, by:
- Prompt engineering
- Model-specific behavior exploitation
- Any creative or technical approach that works in practice

Scope of Work

The freelancer / researcher is expected to:

- Research current LLM behavior (ChatGPT, Gemini, Claude, etc.)
- Propose one or more technical strategies that:
- Increase probability that an LLM attempts to access an external resource
- Or emits a detectable network / resolution signal
- Build a proof of concept (PoC) demonstrating:
- Question rendering (image/text)
- Embedded detection mechanism (URL / QR / encoded link / DNS / beacon)
- Server-side capture of the signal

Provide documentation explaining:
- Why the approach works
- Limitations
- Model compatibility
- Possible countermeasures by AI vendors

Important:
The solution must not require browser extensions, malware, or illegal access to AI providers.

Out of Scope

- Webcam-only proctoring
- Human invigilation
- Browser lockdown tools
- Manual plagiarism detection
- Asking AI providers to cooperate

This is a technical detection & deterrence problem, not policy enforcement.

Deliverables

- Technical Proposal (approach + reasoning)
- Working PoC (code or demo)
- Server-side detection logic
- Documentation / Explanation
- Optional: Comparative analysis across multiple LLMs

This is not a basic web development task — it’s an R&D challenge.

Engagement Model

Paid PoC / milestone-based
Long-term collaboration if successful