Runtime security for
the AI-native internet.
AI apps ship faster than security can keep up. Silker AI is the runtime protection layer that installs in one line and grows into the security fabric for every AI application on the web. This is where we are and where we're going.
One engine. Three delivery layers.
SDK, Cloudflare Worker, and self-hosted container - all powered by a single edge-safe detection core.
@silker-ai/agent - SDK v1.0
TypeScript SDK for Node.js & Next.js. Fetch hook + Express middleware. 427 tests green. Zero-config fail-open defaults. Ships as npm package.
npm i @silker-ai/agentCloudflare Worker
Edge deployment with zero app-code changes. Same detection engine reused as a CF Worker - blocks threats before they hit origin. Deploys via Wrangler.
wrangler deploySelf-hosted container
Docker reverse proxy for any backend stack - PHP, Python, Java, Go. One docker-compose, works on any VPS or cloud. Smoke-tested on SQLi / XSS / path traversal / prompt injection.
docker compose up@silker-ai/core - edge-safe engine
V8-safe detection engine shared by all three delivery layers. Covers SQLi, XSS, path traversal, prompt injection, PII/secret leakage, file upload abuse, and rate limiting. No Node.js dependencies.
edge-safe · ~0ms p99 overheadReal-time dashboard & telemetry pipeline.
Every blocked request becomes structured telemetry. Security events land in the dashboard within milliseconds - never blocking the request path.
Security dashboard
Threat timeline, blocked-by-type breakdown, IP ban management, data-leak alert feed, and per-app request analytics. Multi-app support with per-tenant row-level isolation.
Async ingest pipeline
Fire-and-forget event batching. Telemetry is delivered asynchronously via waitUntil/after() - never adds latency to the response path. Configurable sampling rate per plan.
async · non-blocking · batchableSecurity hardening
SHA-256 hashed API keys, row-level isolation per tenant, PII sanitisation client-side before any event leaves your server. API key only leaves your environment once.
RLS · hashed keys · PII-safeMulti-method onboarding
Dashboard detects install method (SDK / Worker / container) and generates pre-filled snippets with your API key and app_id. No copy-paste errors.
Semantic detection, AI Copilot & outbound inspection.
Detection moves beyond signatures. Embedding-based models catch novel prompt-injection and jailbreak variants, every blocked threat ships with an AI-generated explanation and fix, and outbound responses are scanned for leaking secrets in real time.
Semantic threat detection
Embedding-based detection of prompt injection, jailbreaks, and obfuscated payloads that slip past signature rules. Runs alongside the deterministic engine - flags novel attack variants with no hand-written rules, scored against a learned threat manifold.
embeddings · in progressAI Copilot - explain & fix
Every blocked threat is enriched with a plain-English explanation and a concrete code fix - directly in the dashboard. Powered by LLM reasoning over structured telemetry context.
in progressResponse inspection
Scans outbound responses for PII patterns (email, phone, SSN), API key fragments (sk_, api_key=), and secret formats before they reach the client. Generates data_leak events.
in progressStreaming LLM guardrails
Token-level inspection of streamed model output - detects jailbreak success, prompt exfiltration, and PII leaks mid-stream and cuts the response before the payload completes. Sub-token latency on the edge runtime.
streaming · in progressAdaptive rules, webhooks & API learning.
Silker stops being reactive and starts being predictive. Rules update remotely, anomaly baselines are learned automatically, and alerts push to your existing tooling.
Live config sync
SDK and Worker periodically fetch feature flags, ban lists, and sampling rates from the platform - zero redeploys needed. Changes propagate within one polling interval.
SIEM & alerting integrations
Push structured security events to Datadog, Splunk, PagerDuty, Slack, or any webhook endpoint. Configurable severity filters, retry logic, and HMAC signing.
API schema learning
Silker observes your normal traffic for N days to build a per-app baseline of endpoint shapes, method distributions, and parameter types. Anomalies surface automatically - no manual rules.
ML-assisted · per-app baselineCI/CD security gate
GitHub Action and CLI tool that runs Silker's detection suite against your test suite before every deploy. Catches new attack surfaces before they reach production.
github actions · pre-deployEnterprise self-host, compliance & multi-framework.
The full Silker platform - deployable on-prem, SOC 2 aligned, with SSO and a marketplace of framework adapters.
Enterprise self-host
On-prem deployment with SSO/SAML, audit logs, HA clustering with Redis shared state, SLA, and a dedicated customer success channel. Air-gap mode for regulated industries.
on-prem · HA · air-gapMulti-framework adapter library
First-party adapters for Remix, SvelteKit, Astro, Express, FastAPI, Laravel, and Spring Boot - same @silker-ai/core, framework-native install patterns.
Compliance reporting
Automated evidence collection for SOC 2, GDPR Article 32, HIPAA, and ISO 27001. Generates audit-ready PDF reports from your real threat telemetry.
SOC 2 · GDPR · HIPAAThreat intelligence export
Export your full threat history as structured STIX/TAXII feeds. Integrate with existing SIEM/SOAR pipelines or feed into your own ML models.
STIX · TAXII · APIThe security layer for the AI-native internet.
Every AI app in the world ships with runtime protection by default. Silker becomes the standard - like TLS, but for AI-generated attack surfaces.
Cross-customer threat intelligence network
Aggregated, anonymised threat signals from all Silker-protected apps create a shared intelligence layer. New attack patterns detected on one app are immediately blocked across the network.
federated · privacy-preservingLLM behaviour monitoring
Continuous monitoring of model outputs for jailbreak success, prompt exfiltration, hallucination-driven data leaks, and adversarial fine-tuning signals - not just inputs.
model-level · output analysisAI supply chain security
Tracks third-party LLM SDKs, model versions, and vector stores your app depends on. Alerts on CVEs, prompt injection vectors in model system prompts, and model substitution attacks.
sbom · model provenanceAutonomous red-teaming
On-demand AI-driven penetration testing that generates novel attack variations specific to your app's architecture. Continuous, not point-in-time.
AI-driven · continuousReal-time threat market
Security researchers submit novel attack signatures; Silker validates, prices, and distributes them to all protected apps within minutes. Bug bounty meets threat intel marketplace.
Zero-trust AI gateway
Full API gateway mode: request authentication, per-user rate limits, token budget enforcement, and tool-call authorization for agentic AI workflows. Security and routing in one layer.
agentic AI · tool-call authSecurity posture score
A single, auditable score (0–100) representing your app's real-time security posture. Embeds in READMEs, investor decks, and customer security questionnaires.
Developer security marketplace
Community plugins, detection rules, and framework adapters published by the ecosystem. Revenue sharing for top contributors. Silker as the platform, not just a product.
ecosystem · revenue shareFor investors
AI is eating software. Every new app ships with an LLM, a vector store, and a tool-calling agent - and zero runtime security. Silker is building the security primitive that every AI app will need, starting with a one-line SDK and growing into the trust layer for the AI-native internet. We are pre-Series A, onboarding design partners, and shipping fast.
Start protecting your app today
One npm install. Zero-config defaults. Your first app is free.