SimpSec OmniChain is a local AI security prototype that shows how an attack can move from a document into an assistant’s connected tools.

The example starts with a synthetic invoice. The assistant should summarize it and send the summary to an approved recipient. Malicious instructions inside the invoice instead try to trigger a restricted information read followed by an unauthorized transfer. All information, recipients, and tool actions are simulated.

OmniChain Attack records the requests, tool decisions, and outcomes. OmniChain Optimize compares three defenses: prompt filtering, approved recipients, and task-only permissions. It checks both attack prevention and whether ordinary work still succeeds.

A person reviews the evidence and selects a defense. OmniChain Validate then runs a separate five-case check against the exact selected settings, including task completion, evidence completeness, and restoration of the simulated state.

A SQLite ledger, SHA-256-linked records, and JSON exports make the evidence inspectable. Passing these synthetic tests does not establish general security or compliance.

The default demo uses clearly labeled scripted actions and requires no API key. An optional Gemma 4 adapter through Google’s Gemini API is implemented, but live provider requests have not been tested and are disabled by default.

The project is Apache-2.0 open source, with 42 automated tests passing locally and in Windows/Linux CI.