Thanarak
Kanyaprasit

Red Team / Offensive Security / AI Security
Scroll to play
CHAPTER 02
who, quietly
01 — Who
I look for the cracks
in quiet systems.
Then I write down
what I learned.

By day I find the seams where things weren't meant to give way. In between, I keep the camera rolling on a slower hour.

I'm an offensive security researcher focused on red-team operations and the new attack surface that arrives with large language models — prompt-level exploits, agent chain abuse, model supply chain. I like the part of the job that feels like film: long stretches of patient watching, then one decisive frame.

04 Years adversary simulation
37 CVEs & disclosures shipped
Cumulus clouds collected
Interlude — 00:06 / 00:18

Some attacks look like running.

Forward momentum, low to the grass. The fastest path between two systems is rarely the obvious one.

Up next

Focus —
three rooms, and a fourth.

Red Team. AI Security. Cloud & App. Field Notes.

02 — Focus

Three rooms I keep returning to, and a fourth I share aloud.

02 / 01

Red Team
Operations

Adversary emulation, assumed breach, internal pivots, and physical/social blends. Build playbooks that survive contact with reality.

MITRE ATT&CKC2ADPhishingEDR Bypass
02 / 02

AI
Security

Prompt injection, indirect injection, agent tool abuse, model supply chain, jailbreak research, eval harnesses for safety regressions.

LLM PentestRAGGuardrailsMCPRed Eval
02 / 03

Cloud &
App

Web app pentest, cloud config audits, identity attack paths, container escape. The boring layer between the model and the metal.

AWSK8sOWASPIAMBurp
02 / 04

Field
Notes

Public writeups, open-source harnesses, a small Substack on jailbreak taxonomy. The part of the work that doesn't sit in a vault.

WriteupsOSSTalksWorkshops
Interlude — 00:12 / 00:18

The best ones look like standing still.

Patient watching, listening for the soft click. A camera rolling while nothing seems to happen.

03 — Selected work
Things I've taken apart,
on purpose.
001
Shadow Operator
Multi-stage red-team engagement against a regional fintech: phishing → AD → ESC1 → DA in 11 days.
2025 · CONFIDENTIAL
002
Glasshouse
LLM agent harness exploit: indirect prompt injection via a vendor PDF, exfiltrating tool calls.
2025 · RESEARCH
003
Pale Cumulus
Cloud privilege-escalation chain across IAM, SSM, and a misconfigured KMS grant. Five CVEs filed.
2024 · DISCLOSED
004
Quiet Hours
Internal CTF and training program for a 600-person engineering org. Adversary simulation, by belt rank.
2024 · ONGOING
005
Field Notes
Public writeups on jailbreak taxonomy and a small Python harness for RAG red-evals.
ONGOING · OPEN-SOURCE
04 — Contact
Let's talk about
what could break.

Open to red-team engagements, AI-security research collaborations, and the occasional long walk by a lake.