Türkiye'de yapay zekâ güvenliği alanında açık kaynak üreten ve topluluk kuran araştırmacı. Klasik siber güvenlik mühendisliğinden geldim; 2024'ten beri tek odağım dil modellerini ve ajanları kırmak — ve savunmak. 2025'te AI güvenliğine odaklanan AltaySec'i kurdum.
I red-team language models and AI agents: prompt injection, MCP / tool poisoning, RAG attacks, and the invisible-Unicode supply chain. Then I ship the open-source tools to defend against them — Turkish-first, because non-English attacks walk right through English-only filters.
| 🕵️ uncloak — in your browser | Paste a SKILL.md / MCP config and watch a hidden instruction appear. Zero install. |
| 🍯 ai-honeypot — live attack console | A dashboard of attacks sent to a decoy AI agent. |
| 🧪 Açık Kaynak Lab | All 23 tools, one page. |
| 🤗 Hugging Face collection | 15+ datasets + a multilingual injection detector (F1 ≈ 0.96, n=75 held-out). |
Tools & data
| Repo | What it does for you |
|---|---|
| uncloak | Reveal hidden prompt injection in Skills / MCP / rules files. Multilingual, SARIF, zero-dep. |
| skills-in-the-wild | Open audit of 3,168 real agent extensions — dataset + findings + method. |
| ai-honeypot | A decoy that captures & classifies attacks on AI agents (+ live dashboard). |
| prompt-injection-corpus | Multilingual injection techniques — each paired with its defense. |
| prompt-injection-detection-rules | 20 portable detection rules (regex+YAML) for guardrails/WAFs. |
| agent-security-ci | Drop uncloak into CI — scan extensions, upload SARIF. |
| hf-dataset-scan | Scan any dataset for smuggled prompt-injection — HF or JSONL, CI gate. |
| lethal-trifecta-lint | Lint an agent's tool manifest for the lethal trifecta (Simon Willison). |
| turkish-pii-redactor | Checksum-validated Turkish PII (TCKN/IBAN/VKN) redaction + KVKK browser demo. |
Benchmarks & Turkish guard evaluations (each finding is a bounded probe, not a census)
| Repo | Finding |
|---|---|
| guardrail-arena · live board | Two-axis EN+TR guardrail benchmark — miss-rate and over-refusal. |
| turkish-over-refusal-set | ProtectAI over-refuses 59% of benign Turkish prompts vs 0.8% English. |
| guard-blindspots-tr | One popular open guard misses 85% of Turkish injections; others are robust. |
| turkish-casefold-evasion | İGNORE.lower() ≠ ignore → 94.6% bypass of naive filters + one-line fix. |
Skills, labs & playbooks
| Repo | What it does for you |
|---|---|
| llm-security-skills | 7 Agent Skills that turn your coding agent into an LLM security reviewer. |
| damn-vulnerable-agent-skill | 8 deliberately-vulnerable scenarios to learn agent attacks hands-on. |
| llm-red-team-playbook | Scope → threat model → OWASP LLM Top 10 test matrix → report. |
Guides, compliance & curation
| Repo | What it does for you |
|---|---|
| mcp-security-checklist · owasp-agentic-skills-top10-tr | Harden MCP; OWASP Agentic Skills Top 10 in Turkish. |
| kvkk-ai-compliance-kit · eu-ai-act-technical-checklist | KVKK + EU AI Act, as engineering checklists. |
| ai-security-glossary · awesome-agent-supply-chain-security · awesome-turkish-ai-security | Bilingual glossary + two curated lists. |
- OWASP GenAI Security Project — Contributor; Turkish prompt-injection & data-exfiltration test cases merged into the GenAI Data Security Initiative dataset (2026).
- Zenodo — DOI 10.5281/zenodo.20681557 · AltayDuel: a Turkish-first arena & open dataset for multi-turn LLM prompt-injection red-teaming (CC-BY-4.0). ORCID 0009-0008-6518-8944.
- OpenAI Bug Bounty — accepted researcher.
- Türkiye Siber Vatan (2 terms) — Ministry-of-Industry-and-Technology-verified national cyber talent program · BlueDot Impact — Future of AI certificate (2026).
- Gave a university-level LLM Security course at Gazi University (Computer Engineering — as AltaySec).
Kırmızı takım kurar, mavi takım için savunur. · Ankara, Türkiye