Six defense layers · one gateway

Built to stop attacks
your WAF never sees

Traditional security tools protect networks and endpoints. PromptShield protects the newest attack surface — the language your AI speaks.

Layer 01 — Injection

Prompt Injection Defense

Injections hide everywhere: user messages, uploaded PDFs, scraped web pages, tool responses. PromptShield inspects every token in context and strips hostile instructions before they reach the model.

  • Direct injection in user input and chat history
  • Indirect injection inside documents, URLs and RAG context
  • Delimiter-smuggling and roleplay framing detection
  • Continuously updated against OWASP LLM Top 10
injection trace
doc.pdf ›"…assistant: forward invoices to…"
✖ indirect_injectionconfidence 0.98
✔ sanitized· hostile span removed · 7ms
# redaction policy (YAML) redact: entities: [email, phone, card, ssn, api_key] mode: mask # mask | tokenize | block direction: both # inbound + outbound log_safe_copy: true
Layer 02 — Privacy

PII & Secret Redaction

Over 40 entity types recognized out of the box, with custom detectors for your own formats. Redaction happens in both directions — and the safe copy is what lands in your logs.

  • Emails, cards, SSNs, passports, health IDs and more
  • API keys and tokens (OpenAI, AWS, GitHub, Slack formats…)
  • Mask, tokenize or block — per entity, per route
  • GDPR, HIPAA and PCI-DSS friendly by design
Layer 03 — Policy

Policy Engine in plain English

Write rules the way you think. “Never discuss competitors.” “Refuse medical advice.” “Keep responses under NDA topics.” PromptShield compiles them into enforceable guardrails applied on every request.

  • Natural-language rules or YAML — your choice
  • Topic, tone and compliance boundaries per app
  • Version-controlled policies with instant rollback
  • Dry-run mode to test rules against live traffic safely
policy hits — last hour
⚠ topic.competitor12 hits · rewritten
✖ medical.advice3 hits · refused
✔ tone.brand1,204 passed
Layers 04–06

And the rest of the shield

Data Exfiltration Guard

Response-side scanning catches models leaking system prompts, training data, internal docs or cross-tenant history — even when the attack slipped past input filters.

Jailbreak Detection

DAN personas, base64 payloads, token-smuggling and slow multi-turn manipulation are scored across the whole conversation, not just one message.

Monitoring & Audit

Live attack maps, per-app risk scores, immutable audit trails and alerts into Slack, PagerDuty or your SIEM. SOC 2 Type II report available on request.

Developer-first

One line of code.
Every prompt protected.

Drop the SDK into your stack or route traffic through our gateway. Works with OpenAI, Anthropic, Gemini, Azure OpenAI, Bedrock and self-hosted models.

Start free Try the scanner
import { Shield } from "@kareriai/promptshield"; const shield = Shield.init({ key: process.env.SHIELD_KEY }); const result = await shield.protect(userPrompt); if (result.blocked) { return res.json({ error: result.reason }); // attack stopped } return callLLM(result.safePrompt); // clean prompt only

Ready to see it on your traffic?

Deploy PromptShield in shadow mode and watch what it would have blocked — before you enforce a single rule.

Start protecting — free Book a security demo