Loops HouseContext Firewall

NEAR HACKATHON
Links
Team
1 member- MIOwner
Mikhail Oskola
Overview
AI agents that fill in web forms read the whole page, so card numbers, tax IDs, API keys and one-time codes end up in what the model sees. Context Firewall keeps those values out of the model's context. The agent works with placeholders like [CARD_1]; the real value is typed into the form locally at the moment of the click, and only on a page of the same origin where it was seen. It runs as an MCP server over stdio (tools redact, fill, forget) and as a local HTTP server on 127.0.0.1. Detectors cover cards (Luhn and known IIN), US SSN, ITIN, EIN, IBAN (mod-97), common API keys and tokens (AWS, OpenAI, Anthropic, GitHub, Stripe, Slack, Google, PEM private keys, JWT) and one-time codes. On a synthetic corpus of 300 documents with hard negatives, recall is 95 to 100 percent per category except cards at 82 percent; precision is lower for shape-based rules, and the README lists every number. How NEAR fits in: the design puts a confidential-compute (TEE) classifier on NEAR AI Cloud behind the rules, for what rules cannot describe. In this version it is an interface and a stub that is not connected; the submitted version runs in --local mode only, with deterministic rules and nothing leaving the process. The same team also submits one independent project, Last Server (Games track); the two share no code.