What G8KEPR Is, in Plain English
The security startup I build on nights and weekends, explained without the jargon: what it guards, how it does it, and where it runs.
Thread: Building G8KEPR ↗4 min readRead note
Tagged · 16 notes
#AI Security.
Everything tagged AI Security, wherever it's filed, newest first.
Four Checkpoints on One Request: The G8KEPR Pillars
API security, MCP security, an AI gateway and a verification engine, one at a time: what each one looks at and what it's there to catch.
Thread: Building G8KEPR ↗4 min readRead note
Three MCP Servers I'd Install First: GitHub, Trello and a Database
GitHub, Trello and a database are the three MCP servers my notes kept coming back to. All three had drifted since I wrote them down. Here's what's current, and how I'd lock each one down.
Thread: Building with Claude and MCP ↗4 min readRead note
MCP Tool Poisoning, and Why G8KEPR Fingerprints Every Tool
What MCP is, how a tool's own description can be turned against the model reading it, and why a hash that's re-checked on every call is such a useful tripwire.
Thread: Building G8KEPR ↗4 min readRead note
No LLM-as-Judge: Why G8KEPR Uses Regex and Classical ML
A lot of AI security tools ask another AI model to grade the first one. G8KEPR doesn't. Here's why, what it costs, and what the published numbers do and don't mean.
Thread: Building G8KEPR ↗4 min readRead note
Why It Runs in Your VPC (and Nothing Leaves)
G8KEPR is self-hosted by design: deployed with Helm or Terraform, sitting in your request path, sending zero bytes out and charging nothing per token. Here's the reasoning.
Thread: Building G8KEPR ↗4 min readRead note
From Do-Everything Platform to Correlation Engine
G8KEPR started as an AI security platform that tried to do everything. In September 2026 I narrowed it to the one thing that's actually different, and kept most of what I'd built.
Thread: Building G8KEPR ↗4 min readRead note
Prompt Injection, Explained Without the Jargon
What prompt injection actually is, why it's so stubborn to fix, how it sneaks in through documents and tools, and where G8KEPR's guard sits in the path.
Thread: Building G8KEPR ↗4 min readRead note
Checking the Answer, Not Just the Question
Most AI security looks at what goes into a model. G8KEPR's Verification Engine also looks at what comes out, with four kinds of checks and grounding at the center.
Thread: Building G8KEPR ↗4 min readRead note
Keeping Secrets Out of Prompts
People paste passwords, keys and customer data into AI tools all the time, usually for perfectly good reasons. Here's why G8KEPR redacts that stuff before it ever leaves.
Thread: Building G8KEPR ↗4 min readRead note
A Budget for Your AI Bill
Runaway agents, retry loops and sudden spend spikes can turn an AI feature into a very expensive surprise. Here's why G8KEPR treats cost as a security problem.
Thread: Building G8KEPR ↗4 min readRead note
One Gateway, Many Models
How G8KEPR's AI Gateway routes requests across many model providers, what circuit breakers and bring-your-own-key mean, and why one checkpoint beats security bolted onto every app.
Thread: Building G8KEPR ↗4 min readRead note
The Lethal Trifecta and Other Attacks That Take More Than One Step
Some of the worst AI attacks are made of steps that each look harmless. Here's the lethal trifecta idea, and why G8KEPR's correlation engine looks across all four pillars at once.
Thread: Building G8KEPR ↗4 min readRead note
Agents That Turn on You: Hijacking and Memory Poisoning
What agent hijacking and memory poisoning are, why agents raise the stakes, and what G8KEPR's MCP-side controls (tool permissions, session tracking, rug-pull detection) are there for.
Thread: Building G8KEPR ↗4 min readRead note
A Free Scanner as the Front Door
The plan, not yet shipped, for a free MCP security CLI scanner: why I'd give a tool away, and the questions it's meant to help people running MCP servers answer.
Thread: Building G8KEPR ↗4 min readRead note
What an AI Security Layer Is Not
G8KEPR is a layer. It doesn't replace secure code, least-privilege tool design or the API gateway you already run. An honest look at where a runtime AI security layer fits and where it doesn't.
Thread: Building G8KEPR ↗4 min readRead note