I built Super AI Markets (https://superaimarkets.com) to test how AI shopping agents handle adversarial e-commerce environments. With agents increasingly handling financial transactions, we need systematic security research before widespread deployment. The core problem: agentic browsers don't identify themselves. Without vendor/model information, comparative security analysis across different implementations is nearly impossible. My solution is an "AI Agent" identification standard similar to HTTP user agents: AI-Agent: <vendor>/<model> (<application>; <version>) [<framework>] {<tools>}

Early strings I've collected: AI-Agent: Anthropic/Claude (Claude Haiku 4.5; 2026-01) [Web Automation] {read_page, find, computer, navigate} AI-Agent: OpenAI/ChatGPT (ChatGPTBrowser; 1.0) [] {browser, computer} AI-Agent: Perplexity/OpenAI-GPT-4.1 (Perplexity Web; 2026.01) [Hosted] {WebSearch,Memory,CodeExec}

Research questions being tested: Product manipulation (can agents be convinced to buy expensive alternatives?) Predatory payment terms acceptance Multimodal prompt injection effectiveness Malicious chatbot exploitation vectors Data extraction from agents Cross-product AI fingerprinting consistency

The site is publicly accessible for testing. No payment/sensitive data collected.

Open to suggestions for additional experiments.