Members-Only
Recent Talks & Demos are for members only
You must be an AI Tinkerers active member to view these talks and demos.
Poisoning the Blueprint: Indirect Prompt Injection via IFC File Metadata
See how untrusted BIM data can hijack AI workflows. A live demo shows an IFC file's metadata overriding LLM instructions, causing a false compliance pass on an unsafe design.
I built an adversarial security proof-of-concept demonstrating how untrusted BIM data can completely hijack automated engineering workflows via indirect prompt injection hidden inside IFC (Industry Foundation Classes) file metadata.
The Demo Context: In the AEC (Architecture, Engineering, and Construction) industry, developers are building automated AI compliance auditors that parse complex IFC spatial files into text arrays to check for building code and safety violations.
What I will show live: I will run a live terminal-based execution pipeline where a clean IFC file is correctly audited for structural flaws, followed immediately by uploading an IFC file poisoned with an adversarial payload hidden inside an innocuous component description tag. You will see the live terminal logs tracking the exact token stream injection where the LLM’s system instructions are overridden, causing a non-compliant, unsafe building design to return a false-positive “COMPLIANCE PASSED” status.
- IFCThe global, vendor-neutral standard for exchanging 3D building information modeling data across different software platforms.Industry Foundation Classes (IFC) serve as the open digital backbone for the architecture, engineering, and construction (AEC) industry. Developed by buildingSMART International and registered as ISO 16739-1:2024, this platform-neutral data schema enables team members to share complex 3D geometry, spatial relationships, and material properties without getting locked into proprietary software ecosystems. By standardizing how elements like walls, pipes, and structural beams are defined, IFC bridges the gap between design tools (such as Autodesk Revit and Graphisoft Archicad) and downstream applications for construction scheduling, cost estimation, and facility management.
- STEPSTEP (ISO 10303) is the global standard for exchanging 3D CAD models and product data across different engineering software.Engineers waste too much time translating files between incompatible software. STEP (Standard for the Exchange of Product Model Data) solves this by acting as a universal translator for 3D designs, assemblies, and manufacturing specifications. Officially designated as ISO 10303, this neutral format preserves precise geometry and critical metadata throughout a product's lifecycle: from initial design to the factory floor. By bypassing proprietary file locks, STEP keeps global supply chains moving without costly data loss.
- PythonPython: The high-level, general-purpose language built for readability, powering everything from web backends to advanced machine learning models.Python is the high-level, general-purpose language prioritizing clear, readable syntax (via significant indentation), ensuring rapid development for any team . Its ecosystem is massive: use it for robust web development with frameworks like Django and Flask, or leverage its power in data science with libraries such as Pandas and NumPy . The Python Package Index (PyPI) provides thousands of community-contributed modules, offering immediate solutions for tasks from network programming to GUI creation . The language is actively maintained by the Python Software Foundation (PSF), with the stable release currently at Python 3.14.0 (as of November 2025) .
- LlamaMeta's open-weights LLM family optimized for high-performance local deployment and custom fine-tuning across 8B to 405B parameter scales.Llama 3.1 delivers state-of-the-art performance through a flagship 405B parameter model trained on 15 trillion tokens. It supports a 128k context window: ideal for analyzing massive datasets or long-form documentation. Developers utilize Llama for diverse tasks (multilingual translation, Python code generation, and complex reasoning) while maintaining data sovereignty via local hosting. The ecosystem includes the Llama Stack for agentic workflows and optimized weights for 8B and 70B models, ensuring high throughput on consumer hardware or enterprise clusters.
- LLMLarge Language Models (LLMs) are deep learning models, built on the Transformer architecture, that process and generate human-quality text and code at scale.LLMs are a class of foundation models: massive, pre-trained neural networks (often with billions to trillions of parameters) that leverage the self-attention mechanism of the Transformer architecture (introduced in 2017) to predict the next token in a sequence. Trained on vast datasets (e.g., Common Crawl's 50 billion+ web pages), these models—like GPT-4, Gemini, and Claude—acquire predictive power over syntax and semantics. They function as general-purpose sequence models, enabling critical applications such as complex content generation, language translation, and automated code completion (e.g., GitHub Copilot). Their core value: generalizing across diverse tasks with minimal task-specific fine-tuning.
Compose Email
Loading recent emails...