Skip to content
shadowiq
Use case · Prompt injection defense

Block prompt injection at the gateway, not the post-mortem.

Detection after the fact is a blog post. ShadowIQ blocks, redacts, or escalates suspect prompts inline — and signs every decision for the record.

What this is

Summary

Prompt injection defense in ShadowIQ uses a multi-classifier ensemble (rule-based, small classifier, and LLM-judge quorum) tuned on 2,400+ labeled adversarial samples, enforcing inline at the AI gateway with p99 latency under 75ms and cryptographically signed decisions.

How it fits · explainer

The before / after, in one picture.

PROBLEM · BEFORE SHADOWIQ
Classic detections miss indirect prompt injection via RAG content.
siqSOLUTION · WITH SHADOWIQ
Rule match + small classifier + LLM-judge quorum. 2,400+ labeled adversarial samples, updated weekly.
PILLARS ENGAGEDEvaluateEnforceEvidence
Where it hurts

You've heard this one before.

  • Classic detections miss indirect prompt injection via RAG content.
  • Guardrails that add 300ms latency nobody will ship.
  • False positives that block safe traffic and exhaust ops.
  • No record of what the attacker actually tried.
What we do about it

Three moves.

  1. 1
    Multi-classifier ensemble.

    Rule match + small classifier + LLM-judge quorum. 2,400+ labeled adversarial samples, updated weekly.

  2. 2
    Sub-75 ms enforcement.

    WASM-compiled, parallel-evaluated policies. Latency budget preserved even under burst.

  3. 3
    Full attacker timeline.

    Every suspected injection attempt is signed, chained, and replayable — for your red-team, your IR team, and your auditor.

Outcomes

Numbers, not adjectives.

94%
injection detection recall
< 1%
false-positive rate
74 ms
p99 end-to-end
Frequently asked

Asked, answered, sourced.

Yes. RAG content flows through the same classifier pipeline with context-aware rules. Suspicious retrieved content is flagged or stripped before reaching the model.

Weekly, with major-sample releases on new attack families (e.g., ASCII smuggling, prompt leak via tool descriptions). Customers can subscribe to the RSS of additions.

Yes. Upload a dataset; it becomes a versioned eval you can schedule and share. Your samples stay in your tenant — never train our models.

Ready to see the signet in motion?

Your 30-minute demo. A signed audit trail by the end of it.

We'll wire ShadowIQ into one live workload, block a prompt injection in real time, and hand you a cryptographic receipt — before the meeting ends.