Field notes · red team · AI security
Let's break the agents before advisories do.
I'm Saiprasad - I work in security, and I'm learning AI agent security in public. These are my lab notes: HTB writeups, red-team tooling, and a seven-part field manual on breaking and defending AI agents.
// Latest transmissions
Fresh from the lab notebook.
The newest notes - newest first, no filler. Everything here was broken, built, or both before it was written down.
AI Agent Architecture: The Basics Before You Build or Break One
A security-first guide to agent taxonomy, the execution loop, tools, memory, RAG, MCP, orchestration, and the trust boundaries that matter.
HTB Jugglin
Forela Corporation heavily depends on the utilisation of the Windows Subsystem for Linux (WSL), and currently, threat actors are leveragi...
HTB JingelBell
Torrin is suspected to be an insider threat in Forela. He is believed to have leaked some data and removed certain applications from thei...
HTB Noted
Scenario
Plugout CVE-2024-2879
Scenario
Powershell Code deobfuscation
Scenario
// Field manual - in progress
Building & securing AI agents, in public.
A seven-part field manual on AI agent security - architecture, hands-on labs, threat models, defense, and governance. Written as I learn it, so you can follow along or skip ahead.
Agent architecture basics
What an agent actually is - taxonomy, the execution loop, tools, memory, and the trust boundaries that matter.
Read the note →Before OWASP: Learn the Agent by Building It
Seven hands-on lab steps that teach you the agent by building one - before the frameworks arrive.
In the lab nowOWASP systematization
Mapping what you built onto the OWASP agent threat model.
Defending agents in production
Guardrails, evals, and the controls that survive contact with real users.
MCP security deep dive
The Model Context Protocol under adversarial pressure.
Governing agents
NIST AI RMF, ISO/IEC 42001, the EU AI Act - risk registers and control mapping.
Assuring agents
Control testing, continuous monitoring, cloud deployment, incident response.