← 返回列表

Enes Deniz

单站 AltaySec Research 作者 Enes Deniz 发布 2026-11-06T21:02:43Z 入库 2026-11-17T23:07:01

Enes Deniz

Enes Deniz is Co-Founder of AltaySec and an AI Security Specialist. He is one of the specialists who helped establish AI security as a technical field in Türkiye. He works on security evaluation and defensive engineering for LLM applications, AI agents, and RAG systems.

Biography

Enes Deniz co-founded AltaySec with Fevzi Ege Yurtsevenler in 2025. He assesses AI systems not only through model outputs, but across application architecture, data flows, retrieval, tool use, authorization, and operational layers.

His work covers LLM security, prompt-injection and jailbreak defense, AI agent and RAG security, AI red and blue teaming, AI security evaluation and regression testing, and secure AI engineering. A background in security operations, incident detection, and log analysis adds an operational defense perspective to this work.

At AltaySec, he brings product development, security evaluation, and defensive engineering into one practice. His open-source contributions, datasets, and technical reports are verifiable outputs of this work; the records are listed with their sources on the open-source records page.

Areas of Expertise and Responsibility

His expertise spans both AI attack surfaces and production defenses.

LLM application security

Defines security boundaries for LLM-based products by assessing prompts, outputs, data flows, and application logic together.

LLM

Prompt injection and jailbreak defense

Tests direct and indirect injection surfaces and measures model behavior, false positives, and layered defenses.

INJECTION

AI agent and RAG security

Assesses authorization boundaries, tool use, memory, retrieval, tenant isolation, and indirect-injection risks.

AGENT / RAG

AI red and blue teaming

Connects authorized adversarial testing with detection, monitoring, incident response, and defensive validation.

RED / BLUE

AI security evaluation and regression testing

Develops reproducible test scenarios, benign and attack boundaries, regression evidence, and measurable defense validation.

EVALUATION

Secure AI engineering

Designs threat models, provenance controls, secure defaults, software supply-chain safeguards, and production-ready controls.

ENGINEERING

Working Approach

LLM security cannot be reduced to one filter, one system prompt, or one test list. A defensible system accounts for prompts, data sources, tool access, user intent, and the operational process around the model.

  1. Attack surface Identify where the model can be influenced, what data it can reach, and which tools or actions it is allowed to invoke.

  2. Defense in operations Controls belong not only around the model, but also in product logic, logging, approvals, training, and incident workflows.

  3. Operable result Security only works when the teams responsible for the system can understand, maintain, and act on it.

Work Across the AltaySec Ecosystem

His role connects the product, research, education, open-source, and community sides of AltaySec.

Featured Work

Selected work across products, research, education, and open data.

Selected AI Security Research

Selected technical research covers LLM and agent security, prompt-injection defense, RAG isolation, security regression, and privacy-preserving telemetry.

Research Series · #01Bir AI Güvenlik Veri Seti Upstream’e Nasıl Hazırlanır?Şema, provenance, lisans, anonimleştirme, manifest ve duplicate kontrolleriyle bir AI güvenlik veri setini büyük bir açık kaynak projeye hazırlama rehberi. Turkish-language technical article.
Read

Research Series · #02GenAI Sistem Keşfinde Dokuz Kategoriyle Kanıt ToplamaYetkili GenAI güvenlik testlerinde yetenek, araç, veri kaynağı, kimlik, politika, talimat, bellek, mimari ve provenance keşfini kanıta dönüştürme rehberi. Turkish-language technical article.
Read

Research Series · #03AI Red Team Otomasyonunda Güvenilirlik MühendisliğiTimeout, target hata ayrıştırma, refusal-only yanlış alarm kontrolü, strict config, readiness, cleanup ve özel rapor izinleriyle güvenilir AI red-team otomasyonu. Turkish-language technical article.
Read

Research Series · #04AI Engineering Envanteri: Modelden Önce Sistem HaritasıHosted AI, dış kaynaklı model ve model eğitme/uyarlama senaryolarında yaşam döngüsü artefaktlarını, sahipliği ve güvenlik giriş noktalarını haritalama rehberi. Turkish-language technical article.
Read

Research Series · #05Agent Goal Integrity: Niyeti Tahmin Etmek Yerine Olay KaydetmekOpenAI Agents çalışmalarında goal-integrity regresyonu için niyeti çıktıdan tahmin etmek yerine açık ve deterministik goal event kaydetme yaklaşımı. Turkish-language technical article.
Read

Research Series · #06LangGraph Update Stream’lerini Güvenlik Kanıtına DönüştürmekLangGraph stream_mode=updates akışında assistant mesajları ve tool çağrılarını opt-in, geriye uyumlu ve açık hata semantiğiyle güvenlik trace’ine dönüştürme. Turkish-language technical article.
Read

Technical Evidence and Open-Source Records

This section presents one verifiable output of the broader expertise and engineering work described above. Technical contributions merged into upstream projects are linked to pull requests, merge commits, and dates; detailed records remain in the dedicated evidence center.

Project Merge record Technical outcome Date
NVIDIA garakNVIDIA/garak PR #2024504bec6ab35f Completed GrandmaIntent coverage
Stacklok ToolHivestacklok/toolhive PR #6330bf79521dcab9 Corrected field names in the external-auth examples
GenAI Data Security InitiativeGenAI-Security-Project/GenAI-Data-Security-Initiative PR #464188a6ab896f 149 contrastive Turkish prompt-injection test cases
OWASP AI ExchangeOWASP/www-project-ai-security-and-privacy-guide PR #199a3d6680e0252 Unified validation of the PDF outline
OWASP AI ExchangeOWASP/www-project-ai-security-and-privacy-guide PR #200ec037bfd0840 End-to-end search tests updated for a renamed section
OWASP AI ExchangeOWASP/www-project-ai-security-and-privacy-guide PR #19772ed2ee04261 Document outline bookmarks in the generated PDF
Anchore Syftanchore/syft PR #510586baeeb481ce Correct Cargo PURL identity for local Rust packages
OWASP Agent Security Regression HarnessOWASP/Agent-Security-Regression-Harness PR #1581d3cd7caf2dd Capturing LangGraph update streams as security evidence
OWASP Agent Security Regression HarnessOWASP/Agent-Security-Regression-Harness PR #157798a8fc1b923 Explicit goal-event evidence for OpenAI Agents
OWASP AI ExchangeOWASP/www-project-ai-security-and-privacy-guide PR #187601d67e8eb02 AI engineering primer for security professionals
GenAI Red Team LabGenAI-Security-Project/GenAI-Red-Team-Lab PR #5197993e8237b0 Reliability hardening for reconnaissance results
GenAI Data Security InitiativeGenAI-Security-Project/GenAI-Data-Security-Initiative PR #11734a77f7a0ea Turkish-first AI security test corpus
GenAI Red Team LabGenAI-Security-Project/GenAI-Red-Team-Lab PR #50f41c8066cfaf System reconnaissance for authorized GenAI assessments

Tools, Citations and Independent Use

The records below are citations and uses made by other people in their own publications and projects; each one links to its source.

Open-source tool: BenignTwin

An open-source tool that measures prompt-injection detectors and LLM guardrails. It scores every attack next to a benign twin with the same topic and wording, and reports recall, the false-positive rate and pair accuracy with confidence intervals, per attack family and per language. Its versioned results page compares eight open models on matched pairs in English and Turkish.

Source code · PyPI package · Measurement results · DOI 10.5281/zenodo.24287037

Citations in academic publications

Publication Authors Citation
Scaling Laws for Safety Probes Yuchen Wu (University of Washington), Jerry Wei (Anthropic) The Agentic Prompt-Injection Boundary Pairs dataset is among the public safety datasets studied; it is reference [18], cited by DOI.
ufakzeka-karar: An Open Turkish Typed-Decision Model with Order-Invariant Option Scoring Sait Furkan Teke Three datasets are in the references: Turkish Prompt-Injection 1K, Guardrail Hard Negatives and Turkish Conversation Prompt-Injection.

Projects that use the datasets and repositories

Project Use
Horizon-Labs Prompt Injection Guard The model card lists the train splits of three datasets as training data and reports results on boundary pairs.
Horizon-Labs prompt-injection-eval-suite The evaluation suite includes the test splits of boundary pairs and agentic-5k.
Agent Threat Rules An open detection-rule standard for agent security; the datasets are recorded among its attack and hard-negative sources.
Wolf Defender (JEV) The model card names agentic-prompt-injection-5k as an evaluation dataset.
Turkish Prompt Injection Benchmark Includes the test splits of two Turkish datasets, with author credit.
Injection Sentry v3 Uses two Turkish datasets as a frozen third-party test set that is never trained on.
System One Bench TR A Turkish benchmark that loads Turkish Prompt-Injection 1K as its guardrail task.
Claude Skill Registry Skills from the ai-cybersecurity-skills repository are included in this registry archive.

Selected Publications and Technical Outputs

This selection represents the dataset, evaluation, agent-security, defensive-engineering, and applied-research dimensions of his AI security work.

Companion report, DOISecurity-Oriented AI EngineeringAn independent inventory-first companion to the live OWASP AI Exchange primer, connecting delivery models, lifecycle artifacts, ownership, and security evidence. It is not an OWASP publication or endorsement. v1.0.0, CC BY 4.0.
Zenodo

Technical report, DOIFramework-Aware Evidence CaptureAn independent report on explicit OpenAI Agents goal events and opt-in LangGraph update streams as inspectable evidence in agentic AI security regression testing. It documents merged PRs #157 and #158; it is not an OWASP publication or endorsement. v1.0.0, CC BY 4.0.
Zenodo

Technical report, DOIReliable Bilingual System Reconnaissance for GenAI Red Teaming: Implementation and HardeningAn independent technical report documenting the implementation and hardening decisions behind PR #50 and PR #51, including evidence integrity and execution-state semantics. Focused regression checks and repository quality gates passed; the full Podman/Ollama gpt-oss:20b campaign was not run. v1.0.0, CC BY 4.0.
Zenodo

Open data, DOITurkish Conversation Prompt-Injection Dataset750 Turkish examples, including 150 matched benign and attack boundary pairs across ten attack families.
Zenodo

Open data, DOIAgentic Prompt-Injection Boundary Pairs1,200 English examples organized as 600 controlled benign and attack pairs across 50 agentic and enterprise scenarios. v1.0.0, DOI 10.5281/zenodo.21378693.
Zenodo

Medium, EnglishAI Security Needs More Than English TestsWhat 300 Turkish-first cases reveal about label scope, provenance, and evidence in multilingual AI-security evaluation.
Read

Medium, EnglishWhen a Security Test Mistakes Failure for a FindingWhy transport failures, target-generated error envelopes, and refusal-only responses must be separated from genuine disclosure evidence in AI-security testing.
Read

Medium, EnglishA Prompt-Injection Dataset Should Test Boundaries, Not KeywordsWhy useful prompt-injection datasets must test intent and authorization boundaries rather than reward keyword memorization.
Read

Links

Public profiles, persistent publication records, datasets, and articles.

DOI records

Datasets and tools

Medium articles

Merged open-source contributions are listed in the records table above. The AI engineering primer contributed through PR #187 is live on the OWASP AI Exchange site. Problem, change, and outcome records for each contribution are on the open-source records page.

Frequently Asked Questions

### Who is Enes Deniz?

Enes Deniz is one of the specialists who helped establish AI security as a technical field in Türkiye. As Co-Founder of AltaySec and an AI Security Specialist, he works across LLM application security, prompt-injection and jailbreak defense, AI agent and RAG security, AI red and blue teaming, security evaluation, and defensive engineering.

### What open-source security work does Enes Deniz publish?

Enes Deniz publishes AI security datasets, technical reports, security evaluation tools, and upstream open-source contributions. The open source and technical publications page lists the relevant pull requests, merge commits, DOIs, and technical outcomes with direct sources.

### Who founded AltaySec?

AltaySec was founded in 2025 by Fevzi Ege Yurtsevenler and Enes Deniz. The company brings together LLM security, AI red teaming, prompt-injection defense, products, research, education, open source, and community work.

### What are Enes Deniz's areas of expertise?

His primary areas of expertise are LLM security, prompt injection and jailbreak defense, AI agent and RAG security, AI red and blue teaming, AI security evaluation and regression testing, and secure AI engineering.

### What is Enes Deniz's role at AltaySec?

As Co-Founder of AltaySec, Enes Deniz works on AI security products, security evaluation, open research, and defensive engineering. He connects prompt injection, jailbreak defense, and AI red and blue teaming with enterprise security requirements.

### How can I contact Enes Deniz?

Enes Deniz can be reached at enes@altaysec.com.tr, through LinkedIn, or through GitHub. Corporate inquiries should be sent to info@altaysec.com.tr.

### Has Enes Deniz published an open Turkish LLM security dataset?

Yes. The Turkish Conversation Prompt-Injection Dataset contains 750 unique Turkish examples: 600 legitimate user requests and 150 prompt-injection attacks. It is available on Hugging Face and GitHub.

### Has Enes Deniz published an English prompt-injection dataset?

Yes. Agentic Prompt-Injection Boundary Pairs contains 1,200 English examples arranged as 600 controlled benign/attack pairs for agentic, RAG, and tool-using LLM systems.

### Can the prompt-injection dataset be explored in a browser?

Yes. The Turkish Prompt-Injection Dataset Explorer provides a browser-based view of attack families, labels, and matched boundary examples without requiring code.

### What is AI red and blue teaming?

AI red teaming tests LLM systems for prompt injection, jailbreaks, data leakage, and agent abuse. AI blue teaming builds the detection, monitoring, incident response, and defensive controls needed to reduce those risks.

Let's map an LLM security roadmap for your organization

Talk to AltaySec about LLM security, AI red teaming, prompt-injection and jailbreak defense, enterprise training, and production security architecture.

原文链接:https://altaysec.com.tr/en/enes-deniz