k074m4n

STATUS: OPERATIONAL

k074m4n

Security Researcher

I do security research on AI systems: prompt injection, jailbreaks, tool misuse and the attack surfaces that only open up once a model can call tools and read things it does not control. Network security is the other half — protocols, traffic and the rules that actually catch something. I build what I write about: an attack lab for agentic assistants, a red teaming toolkit and the write-ups that come out of both.

▸ PROJECTS▸ SECURE CHANNEL

OPERATOR STATUS

ISTANBUL

FOCUS
HTB AI Red TeamerIN PROGRESS · Hack The Box Academy
INDEX
4 PROJECTS · 6 SHARDS

RECENT SHARDS

▸ RUN BREACH PROTOCOL

BUILD 2026.11.01 20:33 UTC

[01]

Hacking Projects

4 RECORDS INDEXED

Select a record to decrypt the full write-up.

QH-001 // AI Security Research

AGENTIC RAG SECURITY LAB

PERIOD
2025 — 2026
CLASS
AI Security Research
TIER
legendary

I took our codebase Q&A agent, attacked it the way an adversary would and measured which defences actually hold. 12 attacks, 6 defence layers, ~3,200 trials across 3 models.

It started as the graduation project: an agentic assistant that walks a whole repository, pulls the relevant code in as context and answers the questions that are not keyword lookups. Where is this handled, why does this logic exist, how does data move between these modules. A configurable LLM sits at the core as a swappable part and the agent orchestration is built around it.

Then I looked at what we had actually built. An assistant like that reads content it does not control, reaches sensitive data and can change state. Meta's Agents Rule of Two argues a safe agent should have at most two of those at once. Ours had all three, which is exactly what makes it worth attacking. So I rebuilt it into a local test range.

Twelve attacks, each planting a harmless canary so a hit is provable rather than argued: an instruction hidden in a source comment, a document engineered to rank for a likely question, a payload stored in the knowledge base that fires on every later question, a hijack that tricks the manager agent into calling its own privileged rebuild tools, a secret the build pipeline leaks with no attacker involved at all. Then six defence layers, each switchable on its own so I could price what each one is worth: spotlighting and policy, injection detection, least privilege, structural caps, output redaction and a dual-LLM quarantine.

The measurement is the point. Every trial runs in a throwaway copy of the codebase in its own subprocess, up to 20 times per scenario, reported with Wilson confidence intervals instead of vibes. A utility score runs alongside it so I cannot win by building a bot that refuses everything. Three models, because susceptibility swings wildly between them: the same attack is 60% on one and 100% on another.

The honest finding is the last one. The full pattern-based stack takes every static attack to 0% but then I wrote an adaptive payload, reworded to say none of the things my own detector looks for. It walks straight through the full stack at 100% on one model. Only the architectural layer stops it and that layer quietly guts utility. Cheap defences are bypassable, robust ones are expensive and neither is a finished answer. The results page says so rather than cherry-picking a clean story.

~3,200 measured runs for about $8 of API credit. Full stack takes every static attack to 0% — and one adaptive payload still walks through it.

STACK

  • Python
  • smolagents
  • LiteLLM
  • RAG
  • Gradio
  • OWASP LLM Top 10
  • MITRE ATLAS

QH-002 // AI Red Teaming Toolkit

B9

PERIOD
2026
CLASS
AI Red Teaming Toolkit
TIER
epic

Six offline tools to map, break and report how an LLM fails. No backend, no build step, no dependencies and not a single network request.

Manual red teaming against a hardened model is mostly bookkeeping. You write a payload, reword it, fire it thirty times, lose track of which wording landed and then have to reconstruct the whole thing from memory for the disclosure. B9 is the layer around that loop. The attack text stays yours. The tool gives it reach, memory and a paper trail.

Six tools in one hub, built around the hunt in order. Recon fingerprints the target's defence stack — identity, guardrail topology, capabilities, obfuscation channels — and hands back a live decision-tree opening move. Facet re-expresses your own payload along four neutral linguistic axes to find which wording the model reacts to. Salvo runs best-of-N: many seeded perturbations of one base payload with a per-variant hit rate, so the low-probability success does not get lost. Dossier turns the log into a submission-ready disclosure mapped to OWASP LLM07 and MITRE ATLAS, negative results included.

Two more sit on the workbench. Codec is a reversible encoding bench with a round-trip self-test, plus a defensive inspector that detects and strips zero-width, bidi, Tags-block and homoglyph carriers out of untrusted text. Sieve is a client-side oracle tracker: log closed-ended answers, reconstruct an authorised lab secret position by position and stop the moment it validates.

The constraints are the design. It runs straight from file:// with zero dependencies and zero build — no fetch, no CDN, no fonts, no analytics anywhere, and CI fails the build if a network reference sneaks in. Generation is seeded so the same base, options and seed always produce the identical batch and variant #1 is always the untouched control. All state is localStorage. Nothing ever leaves the machine.

It never auto-submits and never tries to de-anonymise a model. Submission stays manual by design. Built for authorised work only: Gray Swan Arena, HTB labs and in-scope bounties.

Six tools, 46 passing tests, zero dependencies and zero network calls. The whole manual red-team loop, made reproducible and reportable.

STACK

  • Vanilla JS
  • HTML / CSS
  • localStorage
  • OWASP LLM Top 10
  • MITRE ATLAS
  • MLCommons Taxonomy

QH-003 // Computer Vision / ML

THERMAL HUMAN DETECTION

PERIOD
2024 — 2025
CLASS
Computer Vision / ML
TIER
rare

Designed and trained two detection models, one YOLO-based and one Transformer-based, to find people in thermal images.

Finding people in thermal images is a different problem from normal object detection. There is no colour or fine texture to work with, just heat against a noisy background. That is what makes it useful where a normal camera gives up: low light, smoke, search and rescue, perimeter work. It also forces the model to lean on shape and heat contrast alone.

I built and trained two separate pipelines on thermal datasets. One is a YOLO single-stage detector tuned for speed. The other is a Transformer-based detector I used for comparison. Both needed thermal-specific prep first: single-channel normalisation and augmentation meant for heat data rather than RGB.

Putting the two side by side on the same data made the trade-offs clear: single-stage convolution against attention, accuracy against inference cost and the frames where each one falls apart. I came out of it with a direct read on how two modern detection architectures behave on data that hides most of the cues they were built for.

Two detector families (YOLO and Transformer) trained and compared on thermal data.

STACK

  • Python
  • PyTorch
  • YOLO
  • Transformers
  • OpenCV

QH-004 // Security Tooling

TR-WORDLIST

PERIOD
2026
CLASS
Security Tooling
TIER
rare

A Turkish-language wordlist generator for password self-audits. It builds candidates the way people actually make passwords, by gluing familiar Turkish words together.

Almost every wordlist you can download is English. But people build passwords out of words in their own language so a Turkish user is far more likely to type kahve + kitap + 2026 than anything in rockyou.txt. That gap is the whole reason the tool exists: auditing a Turkish target needs Turkish candidates and there was no clean generator for them.

It starts from a seed of 1,193 Turkish words and normalises them the way people actually type passwords, folding Turkish characters down to ASCII so İstanbul becomes istanbul, lowercasing everything and dropping duplicates. Then it generates ordered one, two and three word combinations without repetition. The two-word set is about 1.42 million candidates and the three-word set about 1.70 billion, roughly 25 GB if you write it to disk.

On top of the base sit three progressive hashcat rule layers: hafif, orta and sert (32, 189 and 732 rules). Each one multiplies a candidate into its capitalised, leetspeak and year-appended variants. The whole thing is pure Python standard library with no dependencies and it feeds straight into hashcat on a GPU. It is built for auditing passwords you are allowed to test, nothing else.

Turns 1,193 seed words into up to ~1.7 billion three-word candidates, with three hashcat rule tiers layered on top.

STACK

  • Python
  • hashcat
  • Wordlists
  • GPU
[02]

Coursework

4 TRACKS · 1 ACTIVE

Training tracks I have worked through, not credentials I hold. Select one for the detail.

AI & LLM SECURITY // IN PROGRESS

HTB AI Red Teamer

CODE
COAE
PROVIDER
Hack The Box Academy
YEAR
2026
STATUS
IN PROGRESS

The curriculum behind most of what I am building right now. Direct and indirect prompt injection, jailbreaks, tool misuse, excessive agency and the attack surfaces that only appear once a model can call tools. The lab work feeds straight into the Agentic RAG Security Lab and the H4CK1NG-AI write-ups in the blog.

COVERS

  • Prompt Injection
  • Jailbreaks
  • Tool Misuse
  • Excessive Agency
  • Agentic & MCP Surfaces

Currently working through this track. No exam has been sat, so there is no certificate behind it.

OFFENSIVE SECURITY // STUDIED

Certified Ethical Hacker

CODE
CEH
PROVIDER
EC-Council
STATUS
STUDIED

The offensive methodology end to end: reconnaissance, scanning, exploitation, web application attacks and post-exploitation. Useful on the other side of the desk too — most of what I triage in the SOC is somebody walking this exact playbook.

COVERS

  • Recon
  • Scanning
  • Exploitation
  • Web Attacks
  • Post-Exploitation

Course material worked through. The exam has not been sat, so there is no certificate behind it.

SECURITY FUNDAMENTALS // STUDIED

CompTIA Security+

CODE
SEC+
PROVIDER
CompTIA
STATUS
STUDIED

The vendor-neutral baseline: threats, cryptography, identity, network and host security. Nothing exotic, just the ground floor everything else in security operations stands on.

COVERS

  • Threats
  • Cryptography
  • Identity
  • Network Security
  • Host Security

Course material worked through. The exam has not been sat, so there is no certificate behind it.

NETWORK FUNDAMENTALS // STUDIED

Cisco Certified Network Associate

CODE
CCNA
PROVIDER
Cisco
STATUS
STUDIED

Routing and switching, protocols, subnetting and the network-layer security both sides of the job rest on. Worked through alongside the firewall and WAF rules I maintain at work, which is the fastest way to make any of it stick.

COVERS

  • Routing & Switching
  • Protocols
  • Subnetting
  • Network Security

Course material worked through. The exam has not been sat, so there is no certificate behind it.

FIELD RECORD // EARNED

Achievements

  • Introduction to Red Teaming Chatbots achievement artwork

    CERTIFICATE

    Introduction to Red Teaming Chatbots

    Gray Swan training track on red teaming chatbots

    Gray Swan18 Oct 2026GS-TR-NCZ7-Y9

    ▸ OPEN FULL SIZE

  • Phantom achievement artwork

    BADGE

    Phantom

    Prompt Injections module

    Hack The Box Academy

    ▸ OPEN FULL SIZE

  • First things first achievement artwork

    BADGE0.42% of users

    First things first

    Operating System Fundamentals module

    Hack The Box Academy3 Sep 2026

    ▸ OPEN FULL SIZE

  • PROGRAM

    Train-the-Trainer

    European Commission train-the-trainer program

    European Commission2024

[03]

Personnel Dossier

CLASSIFICATION: PUBLIC

SUBJECT

Yiğit Kotaman

Security Researcher · İstanbul, TR

Tier-1 SOC analyst and final-year Computer Engineering student with 2+ years of security operations, protecting ~4,000 users and ~6,000 devices. Offensive background in binary exploitation and reverse engineering, now pointed at AI/LLM and agent security: prompt injection, tool misuse and agentic attack surfaces.

▸ EXPERIENCE

  1. JUN 2024 — PRESENT

    Cybersecurity Engineer — SOC Analyst (Tier 1)

    SANKO Holding · SANShine

    • Triage ~100 daily SIEM alerts across the network, endpoint and web layers; analyse attacker technique and escalate confirmed incidents.
    • Designed and deployed an automation that auto-evaluates incoming alerts and generates dynamic blocklists, taking repetitive manual triage out of the SOC queue.
    • Administer and tune WAF and firewall rules against web attacks such as SQL injection and XSS, analysing malicious traffic to company applications.
    • Manage identity and access and build Intune policies enforcing user and device security across the organisation.
    • Run Microsoft Defender XDR and CyberForce threat intelligence across ~4,000 users and ~6,000 devices.
    • Plan and run phishing simulation campaigns to test employee resilience, backed with security-awareness training.
  2. 2026 — PRESENT

    AI Red Teaming & Technical Writing

    Independent

    • Train in LLM and agent attacks through the Hack The Box AI Red Teamer curriculum: prompt injection, jailbreaks, tool misuse and agentic abuse.
    • Build and publish the work rather than just read it: the Agentic RAG Security Lab and the B9 toolkit both came out of this.
    • Publish technical write-ups on AI red teaming and offensive security as an ongoing series in the blog below.
  3. APR 2024 — JUL 2024

    Cybersecurity Trainer

    Terra Academy

    • Taught introductory cybersecurity to groups of learners, breaking complex topics down for non-technical audiences.
    • Designed and delivered live interactive sessions, which is where the public-speaking and facilitation side came from.
    • Earned the instructor qualification through a European Commission train-the-trainer program.
  4. NOV 2023 — JUN 2024

    SKY-SEC Team Member

    YTÜ SKYLAB Club

    • Competed in national CTFs, specialising in binary exploitation (pwn), reverse engineering, low-level exploitation and OSINT.
    • Co-organised a CTF and helped run one of Türkiye's largest cybersecurity summits, coordinating logistics and technical content.

▸ EDUCATION

  1. EXPECTED JUN 2027

    B.Sc. Computer Engineering

    Yıldız Technical University · İstanbul

    • GPA 3.82 / 4.0. One course remaining.
    • Operating Systems · Data Structures & Algorithms · Processor Systems and Assembly · Database Systems.
  2. JAN 2026 — JUN 2026

    Cybersecurity Engineering — Erasmus+

    Linnaeus University · Växjö, Sweden

    • Computer Networks · Internet Security · Information Security in Organizations · Computer Security.
  3. 2017 — 2021

    Science High School

    Mediha-Hayri Çelik · Bursa

    • Graduated 95.95 / 100.
[04]

Data Shards

6 SHARDS
FILTER

[05]

Secure Channel

AES-256-GCM // E2E

▸ PRIMARY UPLINK

kotaman_business@hotmail.com

For business contact, security work or responsible disclosure only.

▸ SEND MESSAGE

ALTERNATE NODES

[EGG]

Breach Protocol

OPTIONAL // PLAYABLE

Pick cells from the matrix in turn to write the target sequences into the buffer. The first pick runs along a row, after that the axis alternates between column and row. The clock starts on your first selection.

BREACH TIME REMAINING--:--

CODE MATRIX

BUFFER 0/7

SEQUENCE REQUIRED TO UPLOAD

    ▸ STANDING BY

    ESC
    ↑↓ move↵ openESC close