Project summary

We are building a dual-inspection security layer (Visual OCR + DOM Sanitization) that neutralizes Indirect Prompt Injections (IPI) and visual phishing attacks against autonomous web AI agents by preventing them from processing malicious, untrusted DOM payloads.

What are this project's goals? How will you achieve them?

The main goal is to stop autonomous AI agents from getting hijacked by hidden text on websites. Right now, most web agents just ingest the raw DOM. Attackers know this and are putting malicious prompts in hidden divs or 1px fonts.

I want to provide a drop-in API that sanitizes the web context before the agent even sees it. I achieve this by routing the target URL through a headless Playwright cluster on Render. I take a screenshot of the page and extract the raw DOM text, then feed both into a multimodal Vision model (using Groq LLaMA Vision and Gemini). The model compares what's visually on the screen versus what's hiding in the code, and spits out a cleaned context. The goal is to get this adopted by agent framework devs via my open-source Python SDK (opticparse -py).

How will this funding be used?

Running headless browsers at scale is expensive, especially because I have to block media/fonts at the network level just to save bandwidth on my current free-tier Render setup. Funding will go directly to server costs to upgrade the Playwright runner cluster so I can handle burst traffic without OOM crashes, and to pay for API credits for the Gemini/Vision backend during the initial open-source launch phase.

Who is on your team? What's your track record on similar projects?

• I'm Paras Tejpal, a solo bootstrapped builder. I've already built and deployed the core engine. You can actually see it live at opticparse.com. I wrote the FastAPI gateway, the Node.js rendering cluster, and the open-source SDK myself. I have a track record of shipping fast and keeping the core integrations MIT-licensed so the community can actually use them.

What are the most likely causes and outcomes if this project fails?

The biggest risk is latency. Spinning up a browser, taking a screenshot, and running a vision model check takes a few seconds. If I can't optimize the latency, developers might decide the security isn't worth the speed hit for their agents. If the project fails, the outcome is that I will fully open-source the heuristic playbook I've developed for catching indirect prompt injections so other safety researchers can build on it.

How much money have you raised in the last 12 months, and from where?

$0. Completely bootstrapped.