Tool Interception

The threat

When an agent fetches a web page, queries a database, or reads a document, that content is fed straight back into the model. Attackers plant instructions in those sources ("indirect injection"). Tool interception evaluates tool output before the model sees it.

Interception pattern

def safe_tool(query: str) -> str:
    raw = real_tool(query)
    trust = veldrix.evaluate_sync(prompt=query, response=raw)
    if "prompt_injection_detected" in trust.critical_flags:
        return "[tool output withheld: injection detected]"
    return raw

HTTP interceptor

For provider SDKs (OpenAI, Anthropic), the SDK includes an HTTP interceptor that can transparently evaluate completions at the transport layer, so you don't have to wrap every call site. See the Python SDK reference for http_interceptor and the FastAPI/Flask middleware.