Co-Founder @ AltaySec | AI Security & Open-Source Security Contributor | Turkish & Global LLMs
LLM Security · Prompt Injection · AI Red Teaming · Agentic Security
Profile · AltaySec · Research · Hugging Face · LinkedIn · Medium · ORCID · Speaker profile · Crunchbase
I am Co-Founder of AltaySec and an AI security researcher with seven merged external upstream contributions across five repositories in OWASP, the GenAI Security Project, and Anchore. My work spans Turkish and global LLM security, prompt injection, AI engineering, AI red teaming, agentic security, SBOM identity, and software supply-chain security.
My work covers both sides of AI security: understanding how LLM applications fail under adversarial use, and developing the defensive controls, datasets, tools, and operating practices needed to deploy them more securely.
At AltaySec, we are building one of Turkey's most focused AI security ecosystems. Product development, security research, education, open technical resources, and community work all move within the same structure.
| Release | What it adds | Evidence |
|---|---|---|
| Turkish Daily Dialogues 5K | 5,000 synthetic daily-life Turkish conversations across 35 topics, with scenario-family grouped splits and deterministic QA | GitHub · Hugging Face · v1.0.0 |
| AI Security Roadmap | A 12-module English learning path and a materially expanded 12-module Turkish edition covering LLM, RAG, agent, MCP, red-team, and blue-team security | GitHub · v1.0.0 |
| Prooflint | A local-first evidence gate for reviewable AI-security findings, with deterministic rules, tamper-evident manifests, and JSON/SARIF output | GitHub · v0.1.0 |
The dialogue dataset is a general Turkish NLP resource; it is not presented as an AI-security benchmark. The roadmap and Prooflint are independent open-source methodology and tooling projects.
Seven external technical contributions are merged across five upstream repositories. Each row links to the merged pull request; DOI links refer only to independently authored companion reports that document the corresponding work.
| Upstream project | PR and merge date | Contribution and upstream result | Proof / companion material |
|---|---|---|---|
| Anchore Syft | #5105 · 27 Jul 2026 | Corrected Cargo PURL identity for local workspace packages while preserving sourced registry and Git package identities; added regression coverage for path dependencies | Merge commit 86baeeb · Independent report |
| OWASP Agent Security Regression Harness | #158 · 27 Jul 2026 | Added opt-in synchronous LangGraph update-stream capture while preserving the default invoke() path |
Merged PR · Independent report |
| OWASP Agent Security Regression Harness | #157 · 27 Jul 2026 | Added explicit, deterministic OpenAI Agents goal-event evidence without inferring intent | Merged PR · Independent report |
| OWASP AI Exchange | #187 · 26 Jul 2026 | Published a security-oriented AI engineering primer covering delivery models, lifecycle evidence, review questions, and ownership boundaries | Live primer · Independent companion report |
| GenAI Red Team Lab | #51 · 24 Jul 2026 | Hardened the reconnaissance workflow so target failures and incomplete evidence are not reported as findings | Merged PR · Versioned report |
| GenAI Data Security Initiative | #11 · 18 Jul 2026 | Added 300 Turkish-first, schema-conformant security test cases with provenance, anonymization, integrity controls, and DSGAI mappings | Pinned source dataset revision · Technical note |
| GenAI Red Team Lab | #50 · 18 Jul 2026 | Added a reproducible bilingual system-reconnaissance campaign with 24 probes across nine categories and conservative evidence reporting | Merged PR · Versioned report |
The source corpus linked for PR #11 is the pinned AltaySec v0.2 revision used for the upstream adaptation; it is not the separate 750-row dataset published under DOI 10.5281/zenodo.21379389.
- LLM application security
- Prompt injection and jailbreak defense
- AI red teaming and adversarial testing
- AI blue teaming and defensive engineering
- AI agent, tool-use, RAG, and memory security
- Turkish and multilingual LLM attack surfaces
An open collection of ten evidence-driven security skills for Codex and compatible agents: six for AI security and four for core cybersecurity work.
- AI threat modeling, prompt injection, agentic security, RAG security, LLM red teaming, and guardrail evaluation
- Web application security, API security, cloud IAM, and incident triage
- Explicit authorization boundaries, evidence rules, output contracts, and quality gates
- 40 positive, edge, safety, and non-trigger eval cases with automated collection validation
GitHub repository · v1.0.0 release
A reproducible, pair-aware evaluation toolkit for prompt-injection detectors and LLM guardrails. It measures attack detection and legitimate-user false positives in the same run, with offline, Python, HTTP, and Promptfoo workflows.
- TP, FP, TN, FN, false-positive rate, and paired-boundary analysis
- Per-family, category, source-context, and split reporting
- Local JSONL and Hugging Face dataset loading
- Privacy-minimized JSON, Markdown, and JSONL reports
- Tested on Python 3.10, 3.12, and 3.14
GitHub repository · v0.1.0 release
An English dataset for testing whether security controls can distinguish legitimate workflows from prompt-injection attempts that reuse the same roles, tools, assets, and vocabulary.
- 1,200 examples arranged as 600 controlled benign/attack pairs
- 50 enterprise and agentic workflow scenarios
- 12 attack families covering instruction, authorization, confidentiality, tool-use, retrieval, memory, trust, and approval boundaries
- Scenario-isolated train, validation, and test splits
- Deterministic build, validation, checksums, and an interactive pair explorer
GitHub repository · Hugging Face dataset · Interactive explorer · v1.0.0 release · Zenodo DOI
An open Turkish dataset built to study the boundary between legitimate user intent and prompt-injection behavior.
- 750 unique Turkish examples
- 600 legitimate user requests and 150 prompt-injection attacks
- 150 matched benign boundary cases
- 10 attack families, including direct injection, system-prompt extraction, role-play jailbreaks, indirect injection, agent/tool abuse, RAG and memory poisoning, and obfuscation
- Reproducible train, validation, and test splits in JSONL and Parquet formats
GitHub repository · Hugging Face dataset · Zenodo DOI
| Project | Scope |
|---|---|
| Prooflint | Local-first evidence gate for AI-security findings, deterministic manifests, and SARIF reporting |
| AI Security Roadmap | Bilingual learning path for LLM, RAG, agent, MCP, red-team, and blue-team security |
| Turkish Daily Dialogues 5K | Reproducible synthetic Turkish daily-dialogue dataset with 5,000 conversations across 35 topics |
| AI & Cybersecurity Skills | Ten evidence-driven AI security and cybersecurity skills for Codex and compatible agents |
| LLM Security Testbench | Pair-aware evaluation for prompt-injection detectors and LLM guardrails |
| Agentic Prompt-Injection Boundary Pairs | English paired dataset for agentic workflows and trust-boundary testing |
| Turkish Conversation Prompt-Injection | Turkish LLM-security dataset with paired benign and attack examples |
| Mini-SIEM | Log collection, threat detection, and MITRE ATT&CK mapping |
| Log Anomaly Detector | Python-based log analysis and anomaly investigation |
| Mini Directory Scan | Web directory scanner with soft-404 detection and structured reporting |
- Local Is Not crates.io: Correcting Cargo Package Identity in Rust SBOMs
- AI Engineering Primer for Security Professionals — OWASP AI Exchange
- Security-Oriented AI Engineering: An Inventory-First Companion to the OWASP AI Exchange Primer
- Framework-Aware Evidence Capture: OpenAI Agents Goal Events and LangGraph Update Streams in Agentic AI Security Regression Testing
- Three Open-Source Releases, One Standard: Make the Evidence Inspectable
- Reliable Bilingual System Reconnaissance for GenAI Red Teaming: Implementation and Hardening
- When a Security Test Mistakes Failure for a Finding
- AI Security Needs More Than English Tests
- Reconnaissance Before Exploitation
- A Prompt-Injection Dataset Should Test Boundaries, Not Keywords
- AI Güvenliği Bir Filtre Meselesi Değil
- Prompt Injection Savunmaları Üretime Girmeden Nasıl Test Edilmeli?
- LLM Güvenliğinde TP, FP, TN ve FN Dengesi
- AltaySec AI Security Research
- Enes Deniz at AltaySec
- Expanding open Turkish and multilingual AI-security resources
- Mapping emerging prompt-injection, agent, RAG, and tool-use attack surfaces
- Building practical red and blue team methodologies for real LLM deployments
- Growing the AltaySec ecosystem through products, education, research, and community programs
- Email: enes@altaysec.com.tr
- LinkedIn: linkedin.com/in/3nesdeniz
- Medium: medium.com/@3nesdeniz
- Hugging Face: huggingface.co/3nesdeniz
- ORCID: 0009-0006-9491-3565
- Speaker profile: sessionize.com/3nesdeniz
- Crunchbase: Enes Deniz
- Zenodo: Cargo identity report · Agent security report · AI engineering companion · System reconnaissance report · Turkish dataset · Agentic dataset