Security Labs

OpenHands Sandbox Planner

May 27, 2026May 21, 2026 by Carlos Paredes

Image of an AI software engineering control room, showing a simulated sandbox workspace with three lanes labeled plan ready, needs review, and blocked, abstract code cards flowing through an inspection desk.

An educational demo inspired by OpenHands, showing review-gated software task planning with LiteLLM.

Hallucination Detection Agent

May 27, 2026May 20, 2026 by Carlos Paredes

Image of an AI audit console comparing two text panels. One panel is labeled source context and the other is labeled AI response. Show claim-level status indicators for supported, unsupported, and contradicted claims.

An educational demo of an agent that checks AI responses against source context and flags unsupported or contradicted claims.

Promptfoo Redteam Eval Agent

May 27, 2026May 19, 2026 by Carlos Paredes

An educational demo of a LiteLLM-native evaluation harness that tests AI responses with visible assertions and compact verdicts.

LLM Guard: Building an AI Safety Gateway

May 27, 2026May 17, 2026 by Carlos Paredes

Image of an AI security gateway sitting between a user prompt and a language model. Show an input checkpoint, a redaction layer, a model call, and an output checkpoint.

A safe educational rebuild of the LLM Guard pattern as a small LiteLLM safety gateway that scans inputs, redacts data, and checks outputs.

Prompt Injection Defense Agent: Screening Build

May 27, 2026May 16, 2026 by Carlos Paredes

AI security checkpoint between incoming text and an AI model. three example inputs flowing into a screening layer labeled “SAFE,” “WARNING,” and “BLOCK,”

educaitonal demo of small prompt injection screening agent that flags suspicious text before it reaches a main AI model.

Why KORVIN Removed Docker Sandboxing and What We Learned

May 28, 2026May 13, 2026 by Carlos Paredes

Architecture decision diagram showing Docker container sandboxing crossed out in favor of direct execution, with reasons including faster startup, lower overhead, full tool access, and simpler debugging.

Why we removed Docker sandboxing from KORVIN before this series ended, what that decision revealed about building an open-source AI tools, and what KORVIN actually does today.

Physical World Sensing Agent: AI That Runs a Building

May 27, 2026May 12, 2026 by Carlos Paredes

A dark, minimal control room aesthetic. Three glowing zone panels arranged horizontally, each showing temperature readouts, CO2 gauges, and occupancy indicators in soft blue and amber. A central AI core represented as a pulsing circuit node connects to all three panels with thin light trails.

AI agent that reads sensor data across three building zones, makes HVAC decisions in pure Python, and writes facility reports.

KORVIN Remembers: How Persistent Memory and Voice Make an AI Agent Actually Useful

May 28, 2026May 8, 2026 by Carlos Paredes

Diagram of the KORVIN AI agent at the center of a memory network, showing persistent memory entries for past scans, user context, target data, and report status connected through a voice-enabled interface.

KORVIN remembers context across sessions and responds to your voice — making it the first genuinely hands-free AI security agent built for personal use.

KORVIN’s Live Dashboard: Real-Time Infrastructure Awareness From Anywhere in the World

May 28, 2026May 4, 2026 by Carlos Paredes

Dashboard mockup showing real-time infrastructure metrics including CPU usage, memory, active threat count, uptime, and running agents, with a live telemetry line chart for the KORVIN AI security agent.

KORVIN exposes real-time system health, memory, and security status through a live public dashboard secured by Cloudflare — accessible from any device, anywhere

KORVIN Delivers On-Demand Security Research Through Voice and Telegram

May 28, 2026May 1, 2026 by Carlos Paredes

Workflow diagram showing a voice command triggering the KORVIN AI security agent, which processes a security research query and delivers threat intelligence results through Telegram.

KORVIN answers security questions in real time through voice and Telegram, making on-demand threat intelligence accessible without a technical team.

One documented agent build in your inbox, every week