MidStream inspects a model's output while it is still streaming

What MidStream is

Waiting for a model to finish before checking its output wastes the only moment when intervention is cheap. MidStream does inflight analysis on the stream itself, in Rust, with a WASM build for the browser and a QUIC transport.

What changed

Pull request #102 adds the aidefence-core crate built over shared AIDefence pattern packs. It follows pack driven prompt injection detection covering tool invocation attempts, exfiltration URLs and encoded input.

Get started

For the browser or Node build:

npm install @midstream/wasm

Expected result: the WASM bundle available to analyse a stream as it arrives.

For Rust, add the published crates from crates.io to your Cargo.toml.

Detection packs catch patterns. They are a layer, not a guarantee, and should sit alongside permission limits rather than replace them.

Use it today

Practical case: an agent that can call tools, reading untrusted web content. Input is the model stream; the workflow is scan for injection patterns mid flight; the output is an early stop instead of an executed instruction.

Acceptance test: feed it a page containing an obvious injection string and confirm the safety gate fires before the tool call.

Push it further

Experimental commentary. Inflight analysis makes real time intervention possible — rerouting, halting, or asking for confirmation while generation continues. The limitation is false positives: a gate that fires constantly gets switched off.

Falsifiable test: run a week of normal traffic through the packs and count false positives. If the rate is not near zero, tune before enforcing.