Skip to main content
API · MCP · Gateway · AI — correlated on one plane

Each signal looks harmless. Together they're one AI attack.

A runtime security layer for AI apps — API, MCP, gateway, and model-output checks on one request plane, self-hosted in your own VPC or run by us as a hosted service.

Not a human at a keyboard — an AI agent running a multi-stage campaign: a recon probe, a fuzzing burst, a prompt injection, a poisoned tool. Each move rides below the single-pattern alarms. All four checkpoints publish to one threat event bus, and our Cross-Pillar Correlation Engine scores the co-occurrence — low alone, telling when they line up.

Each checkpoint handles the single-pattern attack it can see. What each one blocks by default is marked below, and the limits are published.

Your keys (BYOK)·Self-hosted: nothing is sent to G8KEPR·Hosted trial: findings are stored on our server (one region)
  • 30-day free trial (extend to 90 days free)
  • No credit card
  • Deploys via Helm
Cross-Pillar Correlation Engineillustrative
Coordinated attack correlated
Coordinated attack — 3 signals, 1 incident
185.220.101.34auto-blocked
GATEWAY security
Cost budget exceeded — request refused
185.220.101.34✓ blocked · 12:04:07
MCP security
Prompt injection in tool call blocked
45.155.204.9✓ blocked · 12:04:05
API security
Path traversal probe blocked
103.75.190.11✓ blocked · 12:04:03
GATEWAY security
Prompt-injection guard refused prompt
91.240.118.4✓ blocked · 12:04:01
Compliance posture

Built against SOC 2, HIPAA, GDPR and EU AI Act controls — readiness assessments available under NDA. No third-party certification yet.

Each figure traces to the held-out evaluation or a committed test

0.786
Held-out detection recall
0.857 precision · 3.12% FPR on 2,580 held-out samples · 0.486 recall (198 of 407) on attacks first published after the training data
26/26
SSRF checks pass
11 static guard/sink · 15 live checks
One platform, four checkpoints

Four doors an
AI attack tries.
One platform
at all of them.

AI lets one attacker hit all four surfaces at once, as a single adaptive campaign. Four checkpoints publishing to one threat event bus is what lets a correlator score four quiet signals as one attack. Each card says what that checkpoint blocks by default.

One Request · Four Checkpoints · One Correlation Plane
Inbound Request → API Security → MCP Security
↳ AI Gateway → Verification Engine → Response Returned
each checkpoint publishes its findings to one threat event bus
One platform, not four
Four checkpoints on the same request, sharing one correlation plane — not four products bolted together.
Every door watched
Findings from all four doors land on one bus, where a correlator scores the ones that occur together. See the limits.
Proof, not adjectives

Watch three vectors collapse into one blocked incident.

Deterministic by default — inline regex plus a trained ML classifier, no LLM-as-judge (the optional LLM tier ships off) — so it's reproducible, costs nothing per token, and can run air-gapped when you self-host it.

Deterministic
$0 / token
Air-gapped
12.38 ms
Median request overhead
Measured on the production node, one client, through authentication. No handler, not under load.
94,000+
Labeled attack samples
Public-sourced, labeled corpus the ML classifier is trained on.
~2,500
API routes (v1)
G8KEPR's own REST API, mounted at start-up.
24
Attack Categories
AI / API / MCP, from the labeled corpus.
correlation-engine · live tail
illustrative
12:04:07.001[MCP]tool output asks the user for a password — credential harvesting
12:04:07.44[API]path traversal on GET /files?name=../../etc/passwd
12:04:07.83[GATEWAY]prompt-injection via homoglyph evasion on /v1/chat
12:04:08.02[CORRELATE]src 203.0.113.44 · 3 vectors · 1.2s window · same actor
12:04:08.02[VERDICT]coordinated attack → 1 compound finding across MCP + API + Gateway
Client blocked across the organization for its next requests
Where it sits in your stack

In-path, in three steps.

Not a dashboard beside your traffic — an enforcement layer inside it. Self-host it in your own VPC to keep your keys and data in your network, or start on the hosted trial.

01

Point traffic through G8KEPR

Drop it in as a gateway or sidecar — installed with Helm (an AWS Terraform module provisions the infrastructure), running alongside the gateway you already use. No application rewrite.

02

It inspects in-path

API calls, MCP tool calls and model responses are checked in the request path. Self-hosted, that happens inside your own network and nothing is sent to G8KEPR; on the hosted trial, traffic passes through our single-region server and findings are stored there.

03

It correlates all four pillars and blocks

Findings from API, MCP and the gateway that come from one client are correlated into one incident, and that client is then blocked across your organization for the next 60 minutes.

Zero-risk evaluation

Self-host it and we never see your data.

Self-hosted, detection runs inside your network, on your infrastructure, on your terms. Try it in production without handing anyone your traffic.

Self-hosted: nothing sent to G8KEPR
Cancel anytime
Helm + AWS Terraform module
No LLM judge by default

Nothing to lose to try it

Four reasons evaluating G8KEPR costs you little.

Self-host it in your VPC
Self-hosted, prompts, responses and secrets stay in your network and nothing is sent to G8KEPR. The hosted trial stores findings on our single-region server.
No credit card to start
30 days, the full platform, cancel anytime. Nothing to expense just to try it.
Self-host or managed
Helm charts and an AWS Terraform module ship with it, and a Terraform provider manages API keys; customer-hosted HA with Patroni is on Enterprise. Or start on the hosted trial.
$0 per token
Regex plus a trained classifier by default, not an LLM judge — no metered inference bill as you scale.
The questions a skeptical buyer asks

Straight answers.

One price per tier, full platform
Correlates where a WAF is blind
Self-hosted: data stays in your VPC
No application rewrite
Deterministic, air-gappable

It’s one published price per tier — not a lighter product. You get the same platform running in production today: the detection cascade, MCP security, the gateway, and the verification engine. You’re early, not buying less.

Your WAF inspects HTTP payloads. It can’t see an MCP tool-poisoning attempt, a prompt-injection evasion, or the fact that those and a path-traversal probe come from the same client in the same window. G8KEPR correlates across all of them — your WAF is one of the fragments it ties together.

It depends on how you run it. Self-hosted (Helm charts, with an AWS Terraform module for the infrastructure), detection runs inside your VPC and nothing is sent to G8KEPR. On the hosted service, including the free trial, the traffic you route through it passes through our single-region server and findings are stored there. Either way the default detectors make no calls to a hosted LLM, so there is no per-token bill.

You don’t have to rip anything out on day one. Run G8KEPR as your gateway or alongside your existing one, point traffic through it, and it starts checking requests — no application rewrite.

No. By default the detection path is inline regex plus a trained ML classifier — no LLM-as-judge; the optional LLM tier ships off — so it’s reproducible, costs $0 per token, and can run air-gapped when self-hosted. The trained models and detection pipeline are ours; the training corpus is public-sourced, and no third-party API makes the call.

Your point solutions are watching three doors. Close all of them with one platform.

Stand up cross-pillar AI security and watch a coordinated attack get correlated into one incident, with the client behind it blocked across your organization.

MCP tools
Tool-call interception
LLM traffic
Provider-level inspection
API gateway
Request / response guarding
One cross-pillar platform
One finding bus across all four pillars
Self-hostable in your VPC
Append-only audit log
  • 30-day free trial (extend to 90 days free)
  • No credit card
  • Hosted trial or self-host
  • Cancel anytime