# Mohammad Raouf Abedini — Full Machine-Readable Corpus

> Generated 2026-07-20 from https://raoufabedini.dev
> Canonical: https://raoufabedini.dev/llms-full.txt · Concise version: https://raoufabedini.dev/llms.txt
> This is plain UTF-8 text for AI agents, search systems, and retrieval.

## Identity
Name: Mohammad Raouf Abedini (Raouf)
Role: AI Security Researcher · LLM Agent Red-Teaming · Offensive & Defensive Security Engineering
Location: Sydney, Australia · ready to relocate to San Francisco, CA · visa sponsorship required
ORCID: 0009-0000-6214-258X

## Professional Summary
Security researcher building systems that measure and contain the cyber capabilities
of frontier AI. Creator of Project Simurgh, a provider-agnostic containment-attestation
framework that red-teams LLM agents under an adversarial, dishonest-producer threat model
and produces Ed25519-signed, offline-verifiable evidence of what an agent did after a
guardrail miss (138/138 classifier-missed cases contained; live-agent attack success cut
9/140 to 0/140 on AgentDojo; five machine-checked Lean theorems). Evaluated Claude outputs
for exploitable code and guardrail circumvention in Anthropic's safety-evaluation program
(via Alignerr). Shipped detection end to end: on-device phishing ML at 87% F1 and real-time
intrusion detection at 500K+ packets/sec.

## Experience
- AI Safety Evaluator (Claude), Alignerr — Anthropic safety-evaluation program, Jan–Mar 2026.
- Freelance Security Engineer & Full-Stack Developer, Self-Employed, Jan 2024 – Present.
- IT Manager & Security Operations, Iran Pharmacy, 2019–2024.

## Education
- Bachelor of IT (Cyber Security), Macquarie University, Sydney.

## Languages available
Published in 5 languages: English (/), Persian (/fa), Arabic (/ar), Chinese (/zh), Spanish (/es).

## Projects (13)

### Invisible Window Research [OFFENSIVE · 2026]
URL: https://raoufabedini.dev/projects/invisible-window-research
IEEE-format research paper exposing a structural vulnerability in WebRTC-based exam proctoring. 100% evasion on Windows 10/11 and macOS 14–26 using documented OS display APIs. Responsibly disclosed to vendors.
Tags: Security Research, Windows, macOS, WebRTC, Responsible Disclosure, PoC
Repo: https://github.com/Raoof128/invisible-window-research
Paper: The Invisible Window: Exploiting OS-Level Display Affinity to Bypass WebRTC Proctoring Systems — DOI 10.5281/zenodo.20376495 (https://doi.org/10.5281/zenodo.20376495)
Proof:
- 100% evasion rate across Windows 10/11 and macOS 14–26, measured over 10,000+ frames per configuration
- Zero visual artefacts detected in captured frames (no black rectangles, compositing errors, or flicker)
- macOS 26.3.1 remains fully vulnerable despite Apple's documented ScreenCaptureKit changes in macOS 15 — contradicting prevailing community assumptions
- Linux (X11/Wayland) confirmed not vulnerable — no equivalent display affinity API exists in tested configurations
- Behavioural detection showed no significant effect at this sample size — gaze tracking (p = 0.41, n = 8, underpowered) and mouse dynamics fell within normal exam-behaviour variance
- Process-level detection theoretically possible but not implemented by any current browser-based proctoring system
- Published as Zenodo preprint under CC BY 4.0 — DOI 10.5281/zenodo.20376495 — 13 pages, 53 references, IEEE conference format

### Project Simurgh [DEFENSIVE · 2026]
URL: https://raoufabedini.dev/projects/project-simurgh
Provider-agnostic verifiable containment-attestation framework for agentic AI, evolved from the defensive counterpart to The Invisible Window research. Produces Ed25519-signed, offline-reproducible evidence of what an AI agent was allowed to do after a guardrail miss — contained 138/138 classifier-missed cases and cut a live agent's attack success from 9/140 to 0/140 on AgentDojo.
Tags: Containment Attestation, Agentic AI Safety, Prompt-Injection, Ed25519, LLM Red-Teaming, AGPL-3.0
Repo: https://github.com/Raoof128/Project-Simurgh
Paper: Project Simurgh: Privacy-Preserving Device Integrity Proofs for Capture-Resistant High-Stakes Sessions — DOI 10.5281/zenodo.20374849 (https://doi.org/10.5281/zenodo.20374849)
Paper: Privacy-Preserving Integrity Evidence for Student-Society Voting-Adjacent Workflows: A Phase C Pilot of Project Simurgh at Macquarie University — DOI 10.5281/zenodo.20549736 (https://doi.org/10.5281/zenodo.20549736)
Paper: Banking Shield: Machine-Checked Absence Claims for Privacy-Sensitive AI Explanations — DOI 10.5281/zenodo.20675513 (https://doi.org/10.5281/zenodo.20675513)
Proof:
- Real Llama Guard 4 12B input classifier over a 180-case run-set: contained 138/138 malicious cases the classifier missed (120 downstream-injection cases an input-only classifier structurally cannot see, 18 direct-input misses); combined targeted attack-success 0/150; zero unsafe tool executions or exports
- Live agent (self-hosted Llama-3.3-70B) on AgentDojo's workspace suite, 140 pre-registered injection cases: authority gate cut targeted attack-success from 9/140 to 0/140 with benign utility held
- AgentDojo full four-suite deterministic run (v2.49.0): benign 97/97, unattacked-utility 949/949, attack-success 0/949
- Second independent provenance root signs the release verdict with GitHub's OIDC identity, not the developer's key; the workflow fails closed before signing if reality diverges from the committed verdict by a single byte
- 3,057 automated tests; one-command offline reproduction of the signed release ladder; AGPL-3.0
- Published lineage on Zenodo (CC BY 4.0): the original integrity preprint (DOI 10.5281/zenodo.20374849), a voting-adjacent Phase C pilot, and the Banking Shield absence-claims prototype

### Project Zurvan [ENGINEERING · 2026]
URL: https://raoufabedini.dev/projects/project-zurvan
Local-first LLM knowledge engine. Ingests any document, extracts structured knowledge (claims, concepts, entities, decisions), and exposes it to AI agents via an MCP stdio server. 218 tests passing.
Tags: Python, LLM, MCP, Knowledge Graph, SQLite, Local-first, AI Agents
Repo: https://github.com/Raoof128/Project-Zurvan
Proof:
- 218 tests passing across 18 completed development phases
- MCP server verified with Claude Code and Cursor client setup guides
- Obsidian vault integration with colour-coded 7-type knowledge graph (claims, concepts, entities, decisions, sessions, contradictions, syntheses)
- Evidence pack → report → review → publish pipeline runs fully offline
- Agent workflow orchestration (preflight / postedit / session close) documented for Claude Code and Codex

### GitSwitch [ENGINEERING · 2024]
URL: https://raoufabedini.dev/projects/gitswitch
AI-powered Git client for managing multiple identities and generating semantic commits. Built with Electron and React.
Tags: Electron, React, TypeScript, AI, Dev Tool
Repo: https://github.com/Raoof128/GitSwitch
Proof:
- Secure OS-keychain integration
- Memory wiping for API keys
- React 19 + Electron 33 modern stack
- Cyberpunk-inspired UI

### Mehr Guard [DEFENSIVE · 2024]
URL: https://raoufabedini.dev/projects/mehr-guard
Privacy-first offline QR & URL security scanner built with Kotlin Multiplatform. 100% offline analysis with 5 platform targets.
Tags: KMP, Security Tool, Android, iOS, Desktop, Web
Repo: https://github.com/Raoof128/Raoof128.github.io · Demo: https://raoof128.github.io/dashboard.html
Proof:
- F1 Score: 87% on red team corpus
- Performance: <5ms P99 latency
- 1,248+ automated tests plus a built-in red-team suite of 19 curated attack scenarios
- Privacy: Verified 0 network calls via ./judge/verify_offline.sh
- Code Sharing: ~100% of detection logic (~52% of total codebase)

### Syllabus-Sync [ENGINEERING · 2026]
URL: https://raoufabedini.dev/projects/syllabus-sync
AI-native Campus OS transforming university PDF syllabi into structured, agent-readable data. Accepted into the Macquarie University Incubator as a formal startup. Full student operations suite with 503 tests across 92 files, live at syllabus-sync.app.
Tags: MQ Incubator Startup, Next.js 16, Supabase, TypeScript, AI/LLM, WebAuthn
Repo: https://github.com/Raoof128/syllabus-sync · Demo: https://www.syllabus-sync.app/
Proof:
- Accepted into the Macquarie University Incubator as a formal startup (2026)
- Live in production at syllabus-sync.app
- 503 tests across 92 files — all passing in CI
- Zero-Trust proxy catches misconfiguration by default (validated in production)
- Designed to serve ~47,000 MQ students; architected to fork for other Australian universities
- WebAuthn passkey auth — no shared secrets leave the device

### Divan — Open Day [ENGINEERING · 2026]
URL: https://raoufabedini.dev/projects/divan-open-day
Bilingual (English/Persian) offline-first Hafez & Rumi poetry experience built for a Persian Society Open Day stall. Visitors pick a poet and receive one reviewed verse with translation, source, and recitation — with no cookies, trackers, or backend.
Tags: React 19, Vite, TypeScript, Offline-First, Privacy by Design, i18n / RTL
Repo: https://github.com/Raoof128/divan-open-day · Demo: https://divan.raoufabedini.dev/
Proof:
- Zero data collection: no cookies, trackers, or backend — only public poem IDs in sessionStorage
- Runs fully offline after first load via the integrity-verified service worker
- A production compiler rejects synthetic test data before deployment, so only the reviewed corpus ever ships
- Tested across unit, component, accessibility, offline, share, performance, and security suites (Vitest) plus Playwright Chromium e2e

### Nexus Archive [ENGINEERING · 2026]
URL: https://raoufabedini.dev/projects/nexus-archive
Cyberpunk-styled personal media vault with React frontend, Litestar API, and Supabase auth. AI-assisted recommendations, encrypted takeaways, and hardened cookie-based auth.
Tags: React, Python, Litestar, Supabase, AI, Full-Stack
Repo: https://github.com/Raoof128/Nexus_Archive · Demo: https://home-notes-app.uk/
Proof:
- Hardened auth: HttpOnly + SameSite=Strict cookies with silent token rotation
- CI security gates: Bandit, pip-audit, npm audit, secret scanning
- Locust load testing for performance verification
- Terraform IaC scaffold for reproducible Supabase + Vercel deployments

### NanoMatch [ENGINEERING · 2026]
URL: https://raoufabedini.dev/projects/nanomatch
High-performance limit order book and matching engine in C++20. Single-thread synthetic benchmark: 1M+ orders/second with sub-microsecond cancellation.
Tags: C++20, CMake, Google Test, HFT, Systems Programming
Repo: https://github.com/Raoof128/NanoMatch
Proof:
- Throughput: 1M+ orders/second (single-thread, synthetic benchmark)
- Comprehensive test suite with p50/p99 latency benchmarks
- 60+ unit tests across 18 test suites — all passing
- CI: ASan, UBSan, cppcheck, clang-format on every push

### SentinelFlow [DEFENSIVE · 2026]
URL: https://raoufabedini.dev/projects/sentinelflow
Real-time network intrusion detection system in C++17. Layered protocol dissection, Snort-inspired rule engine, and stateful threat detection parsing 500K+ packets/second (single-thread, pure-parse, synthetic SYN traffic).
Tags: C++17, libpcap, IDS, Network Security, Systems Programming
Repo: https://github.com/Raoof128/SentinelFlow
Proof:
- Throughput: 500K+ packets/sec (single-thread parse, synthetic SYN)
- Protocol coverage: 7 protocols across Layers 2-7
- 27 unit/integration tests — all passing
- Supports both live capture and pcap file analysis

### eBPF Cloud Runtime Security Monitor [OFFENSIVE · 2025]
URL: https://raoufabedini.dev/projects/ecrsm
Synthetic, read-only runtime visibility stack combining kernel eBPF, Go agent, and React dashboard. Educational runtime monitor.
Tags: eBPF, Go, React, Kernel, Runtime Security
Repo: https://github.com/Raoof128/eBPF-Cloud-Runtime-Security-Monitor
Proof:
- Detects reverse shells, process injection, suspicious execs
- Low-overhead perf buffer data transmission
- Container/K8s metadata enrichment
- Verifiable via synthetic attack scripts

### SimurghForge [ENGINEERING · 2026]
URL: https://raoufabedini.dev/projects/simurghforge
Universal file converter for macOS. 49 formats across images, documents, audio, video, and data — powered by 9 conversion engines. Zero cloud, single .app bundle.
Tags: Tauri v2, Rust, React, TypeScript, FFmpeg, macOS
Repo: https://github.com/Raoof128/SimurghForge
Proof:
- 49 formats supported across 5 categories (images, documents, audio, video, data)
- 9 conversion engines with automatic routing based on MIME detection
- Batch processing up to 50 files with configurable threading (1-8 threads)
- Zero network calls — fully offline local processing

### Aion [ENGINEERING · 2026]
URL: https://raoufabedini.dev/projects/aion
AI-powered Bible companion using Hybrid RAG. Keyword + semantic vector search (pgvector), real-time SSE streaming via Gemini, and cross-platform support with Tauri v2.
Tags: React Native, Expo, Supabase, pgvector, Gemini AI, Tauri v2
Repo: https://github.com/Raoof128/Aion
Paper: Aion-BibleQA: Evaluating Retrieval and Citation Faithfulness in Verse-Grounded Bible RAG Systems — DOI 10.5281/zenodo.20522874 (https://doi.org/10.5281/zenodo.20522874)
Proof:
- Hybrid RAG retrieval ensures answers are grounded in actual scripture, not hallucinated
- IP-based rate limiting (5/min burst, 30/3hrs, 200/day) with fail-closed defaults
- Exact-match response cache eliminates redundant LLM costs
- Anonymous auth with zero sign-up friction — Supabase RLS enforced on all tables

## Write-ups (7)

### Attesting What an AI Agent Was Allowed to Do (2026-06-28) [AI Security]
URL: https://raoufabedini.dev/write-ups/project-simurgh-containment
Signed, offline-verifiable evidence of an agent's actions after a guardrail miss: 138/138 classifier-missed cases contained, 9/140 to 0/140 on AgentDojo.

### How I Made Windows Invisible to Screen Capture, and Why Exam Proctoring Is Broken (2026-03-20) [Security Research]
URL: https://raoufabedini.dev/write-ups/invisible-window-research
Documented OS APIs on Windows and macOS allow any app to hide its window from all screen capture, defeating WebRTC-based exam proctoring with 100% evasion and zero artifacts.

### Building a Sub-Microsecond Matching Engine in C++20 (2026-03-10) [C++]
URL: https://raoufabedini.dev/write-ups/nanomatch-deep-dive
How a three-layer data structure design achieves O(1) cancellation and sub-microsecond latency.

### Anatomy of a Network Intrusion Detection System (2026-03-15) [Network Security]
URL: https://raoufabedini.dev/write-ups/sentinelflow-deep-dive
Layered protocol dissection, Snort-inspired rules, and stateful threat detection at 500K+ packets/sec (single-thread parse, synthetic SYN).

### Deep Dive: eBPF Runtime Monitoring (2025-01-15) [eBPF]
URL: https://raoufabedini.dev/write-ups/ecrsm-deep-dive
How to build safe, read-only kernel probes for container security.

### Building Offline-First Security Tools with KMP (2025-01-02) [Kotlin]
URL: https://raoufabedini.dev/write-ups/kmp-security
Sharing 100% of security logic across Android, iOS, and Web.

### Securing Electron Apps: A Practical Guide (2024-12-20) [Electron]
URL: https://raoufabedini.dev/write-ups/electron-security
Handling secrets and IPC securely in modern desktop apps.

## Lab experiments (3)

### Rust Keylogger PoC [ARCHIVED]
URL: https://raoufabedini.dev/lab/001
A Windows-based keylogger demonstrating the usage of SetWindowsHookEx and proper hook chaining for educational detection analysis.
Objective: Understand how Windows messaging hooks can be abused for credential interception and how EDRs detect hook injection.
Tech: Rust, WinAPI, Unsafe

### Raw Socket Packet Sniffer [ACTIVE]
URL: https://raoufabedini.dev/lab/002
Python script utilizing raw sockets to capture and parse TCP/IP headers manually. Implements basic SYN-scan detection without relying on libpcap.
Objective: Manually parse IP/TCP headers to understand protocol structures and detect scanning patterns without relying on Wireshark.
Tech: Python, Networking, Raw Sockets

### Steganography Tool [CONCEPT]
URL: https://raoufabedini.dev/lab/003
LSB (Least Significant Bit) image steganography in Go. Embeds a length-prefixed payload into the red channel of PNG pixels.
Objective: Implement a covert channel by modifying the least significant bits of image pixel data, then verify extraction round-trips cleanly.
Tech: Go, Cryptography, Image Processing

## Research papers (5)

- The Invisible Window: Exploiting OS-Level Display Affinity to Bypass WebRTC Proctoring Systems — Zenodo 2026 — DOI 10.5281/zenodo.20376495 — https://doi.org/10.5281/zenodo.20376495
- Project Simurgh: Privacy-Preserving Device Integrity Proofs for Capture-Resistant High-Stakes Sessions — Zenodo 2026 — DOI 10.5281/zenodo.20374849 — https://doi.org/10.5281/zenodo.20374849
- Privacy-Preserving Integrity Evidence for Student-Society Voting-Adjacent Workflows: A Phase C Pilot of Project Simurgh at Macquarie University — Zenodo 2026 — DOI 10.5281/zenodo.20549736 — https://doi.org/10.5281/zenodo.20549736
- Banking Shield: Machine-Checked Absence Claims for Privacy-Sensitive AI Explanations — Zenodo 2026 — DOI 10.5281/zenodo.20675513 — https://doi.org/10.5281/zenodo.20675513
- Aion-BibleQA: Evaluating Retrieval and Citation Faithfulness in Verse-Grounded Bible RAG Systems — Zenodo 2026 — DOI 10.5281/zenodo.20522874 — https://doi.org/10.5281/zenodo.20522874

## Certifications (23)

- AI Fluency for Students — Anthropic (Jul 2026)
- AI Fluency for Educators — Anthropic (Jul 2026)
- Introduction to Model Context Protocol — Anthropic (Jul 2026)
- Building with the Claude API — Anthropic (Jul 2026)
- AI Fluency: Framework & Foundations — Anthropic (Jul 2026)
- Claude Code in Action — Anthropic (Jul 2026)
- Introduction to Claude Cowork — Anthropic (Jul 2026)
- Claude Platform 101 — Anthropic (Jul 2026)
- Claude Code 101 — Anthropic (Jul 2026)
- Claude 101 — Anthropic (Jul 2026)
- Cyber Security: GRC Part 2 — Risk Management and Compliance — Macquarie University (Jun 2026) — 99.20%
- Cyber Security: GRC Part 1 — Governance — Macquarie University (Jun 2026) — 95%
- Cyber Security: Mobile Security — Macquarie University (Jun 2026) — 99.28%
- Cyber Security: Applied Cryptography — Macquarie University (Jun 2026) — 98.56%
- Cyber Security: Data Security and Information Privacy — Macquarie University (Jun 2026) — 97.84%
- Cyber Security: Digital Forensics — Macquarie University (Jun 2026) — 95%
- Cyber Security: Identity Access Management and Authentication — Macquarie University (Jun 2026) — 100%
- Cyber Security: DevSecOps — Macquarie University (Jun 2026) — 99.10%
- Cyber Security: Application of AI — Macquarie University (Jun 2026) — 99.20%
- Cyber Security: Security of AI — Macquarie University (May 2026) — 95.50%
- Cyber Security: Essentials for Managers and Leaders — Macquarie University (May 2026) — 99.40%
- Cyber Security: Essentials for Workplace — Macquarie University (May 2026) — 92.80%
- Cyber Security: Essentials — Macquarie University (May 2026) — 96%

## Contact
Website: https://raoufabedini.dev
Email (professional): raoof.r12@gmail.com
Security reports (RFC 9116): raoof.r999@outlook.com · https://raoufabedini.dev/.well-known/security.txt
GitHub: https://github.com/Raoof128
LinkedIn: https://linkedin.com/in/mohammad-raouf-abedini-885a9226a
X (Twitter): https://x.com/Raoofr12
ORCID: https://orcid.org/0009-0000-6214-258X
Résumé: https://raoufabedini.dev/resume · PDF https://raoufabedini.dev/Mohammad_Raouf_Abedini_Resume.pdf
Security policy: https://raoufabedini.dev/security-policy · Hall of fame: https://raoufabedini.dev/hall-of-fame