Show Notes
Hosts: Miriah Peterson, Matt Sharp, Chris Brousseau Recorded: April 2026 Status: Released Most AI systems today are designed to be helpful — not secure. In this episode, we break down how AI systems actually get exploited in production: a real supply chain attack on a widely used AI dependency prompt injection and why it still works image-based (multimodal) exploits tool and agent abuse If you’re building AI — especially at a startup — you are the security team. A widely used AI dependency was compromised via a malicious .pth file: executes automatically when Python starts no import required targets credentials, SSH keys, and environment variables 👉 Just installing the package was enough. This highlights a critical reality: Your AI system is only as secure as your dependencies. Models cannot distinguish between instructions and data External content can override system behavior Still one of the most common AI vulnerabilities 🔗 https://learnprompting.org/docs/prompt_hacking/injection Hidden instructions embedded in images AI interprets images differently than humans Expands the attack surface significantly 🔗 https://arxiv.org/abs/2306.11698 AI systems can take real-world actions via tools Prompt injection → API calls, data leaks, unintended execution Agents amplify risk through autonomy and retries If you’re building AI systems today: separate instructions from data limit tool permissions treat outputs as untrusted validate everything before execution AI systems have an internet-sized attack surface Supply chain attacks bypass all AI safeguards Prompt injection is a fundamental problem AI doesn’t fail safely — it fails wherever your system is weakest LiteLLM incident: https://github.com/BerriAI/litellm/issues/24512 Attack breakdown: https://futuresearch.ai/blog/litellm-pypi-supply-chain-attack/ LLM attack techniques: https://llm-attacks.org/ OWASP LLM Top 10: https://owasp.org/www-project-top-10-for-large-language-model-applications/ Gandalf challenge: https://gandalf.lakera.ai/ We’ve launched a Patreon for Domesticating AI 🎉 Get: early access to episodes behind-the-scenes content bloopers and uncut moments 👉 https://patreon.com/DomesticatingAIPodcast 🎥 YouTube: https://youtu.be/HTTxE7Y1sko What’s the weirdest way an AI system has broken for you? Keep your AI on a leash.
More from Domesticating AI
AI agents can consume more information than ever. That doesn't mean they understand your data. For the Season 1 finale of Domesticating AI, we're joined by Joe Reis, co-author of Fundamentals of Data Engineering, for a conversation that went somewhere completely different than we planned. We start with Joe's practical take on local AI. If subscriptions to frontier models are cheaper than buying hardware, when does self-hosting actually make sense? Privacy, IP, fine-tuning, and inference costs at scale all change that calculation. Then we get into data modeling. Humans are remarkably good at working around bad schemas, missing documentation, tribal knowledge, and ambiguous business definitions. Agents aren't. If machines are going to operate on our data, the structure, relationships, provenance, and meaning of that data become much harder to ignore. From there, things get weird. We talk about corporate knowledge bases, ontologies, the post-literate engineer, AI-generated summaries, whether engineers still need to read the research, and the increasingly shaky state of reproducible AI benchmarks. Topics include: • Local models vs. cloud AI economics • Bonsai 27B and 1-bit models on phones • Data modeling for AI agents • Knowledge bases and organizational knowledge • Ontologies and semantic modeling • The post-literate engineer • AI summaries vs. actually reading research • Reproducibility and AI benchmarks • Why agent security doesn't require Skynet to be dangerous Links mentioned: Joe Reis — The Post-Literate Engineer https://joereis.substack.com/p/the-post-literate-engineer Joe Reis — Practical Data Modeling / Mixed Model Arts https://practicaldatamodeling.substack.com/ The Mixed Model Arts Manifesto https://practicaldatamodeling.substack.com/p/the-mixed-model-arts-manifesto Fundamentals of Data Engineering — Joe Reis & Matt Housley https://www.oreilly.com/library/view/fundamentals-of-data/9781098108298/ PrismML — Bonsai 27B https://prismml.com/news/bonsai-27b DeepSeek V4 Flash https://api-docs.deepseek.com/updates/ William Kent — Data and Reality https://bkent.net/Doc/darxrp.htm Find Domesticating AI: YouTube https://www.youtube.com/@DomesticatingAI Apple Podcasts https://podcasts.apple.com/us/podcast/domesticating-ai/id1873338950 Spotify https://open.spotify.com/show/2WsAR4fvcXzp3vVZGVlkE2 Patreon https://patreon.com/DomesticatingAIPodcast That's a wrap on Season 1. Subscribe so you're here when Season 2 drops. Keep your AI on a leash.
Your side project starts on the free tier. Then it gets interesting. This week, Tod from Catalyst Community joins us to talk about why the homelab is one of the best places to build open-source software, experiment with AI, and learn the infrastructure underneath the tools we use every day. We get into why Kubernetes does not need to turn you into a Kubernetes administrator, how Foundry is making it easier to stand up K3s on your own hardware, and why tools like Helm make adding databases, monitoring, storage, and other services much less painful once the foundation exists. Tod walks us through his own homelab: Proxmox, Debian, a three-node Kubernetes cluster, separate storage and GPU machines, and the open-source tooling he uses to build on top of it. We also get into Tailscale, KServe, Ray, CloudNativePG, Prometheus, Grafana, Loki, Whisper, and running GPUs both inside and outside Kubernetes. But the bigger conversation is about building. Matt talks about the GPU machine he originally bought to train a StarCraft bot and how that experimentation turned into skills he could use professionally. Miriah talks about replacing a Supabase free-tier database with Postgres on her homelab. Tod talks about replacing GitHub-hosted CI/CD infrastructure with his own open-source system. The common thread: doing the foundational work once gives you a platform where the next idea is much cheaper to try. And with modern hardware, open-source software, and LLM-assisted development, the barrier to building something useful at home has never been lower. So our question for you: What are you building for yourself, for fun, or for open source? Projects mentioned in the episode include Foundry, K3s, Kubernetes, Helm, Proxmox, Tailscale, KServe, Ray, CloudNativePG, Prometheus, Grafana, Loki, MLflow, Whisper, Render, Supabase, and Anthropic Claude. Keep your AI on a leash.
Everyone has a list of open-source AI projects they want to try. But installing repos is not the same as building a stack. Ryan Russon joins Miriah, Matt, and Chris to talk through the projects that actually teach you how AI homelabs work: Unsloth for training, Ollama and llama.cpp for local inference, Open WebUI for a usable front door, K3s for hosting, Longhorn and SeaweedFS for storage, and Prometheus/Grafana for observability. We also get into build-vs-buy decisions, speech models like Whisper.cpp, Parakeet, and Granite, why state is where homelabs stop being cute, and why the best open-source projects are the ones that teach you what breaks next. Links mentioned: https://github.com/unslothai/unsloth https://ollama.com/ https://github.com/ggml-org/llama.cpp https://github.com/open-webui/open-webui https://k3s.io/ https://longhorn.io/ https://github.com/seaweedfs/seaweedfs https://prometheus.io/ https://grafana.com/ https://github.com/Arize-ai/phoenix https://github.com/langfuse/langfuse Support the show: https://patreon.com/DomesticatingAIPodcast Keep your AI on a leash.
What happens when the cost of running AI is no longer hidden behind a cloud bill—but something you can physically feel? In this episode of Domesticating AI, Kat Tomanek and Alex Kauffman of Squeez Labs introduce CrankGPT, a fully offline voice assistant that runs on a Raspberry Pi and is powered by a hand crank. As the language model performs inference and generates speech, the crank becomes harder to turn, converting an abstract computational cost into physical exertion. CrankGPT began as an experiment: could Squeez Labs build a responsive, completely self-contained voice AI system without a cloud connection, wall outlet, or battery? Answering that question required much more than placing a small language model on a Raspberry Pi. Kat and Alex had to engineer the complete system, including speech recognition, voice activity detection, language-model inference, text-to-speech, boot time, memory usage, power delivery, and conversational latency. The result is a modular voice stack that uses specialized components rather than one enormous multimodal model. That architecture allowed the team to select each model according to the actual requirements of the device. The voice only needs to be understandable. The language model needs to respond quickly enough for conversation. The speech recognizer needs to run efficiently on a CPU. Anything beyond those requirements consumes resources without necessarily improving the experience. We discuss why local AI is not simply about buying an expensive GPU, why developers should begin with the machines they already own, and how constraints reveal the hardware characteristics that actually matter. In CrankGPT’s case, memory bandwidth and cold-start behavior created more significant limitations than several of the specifications the team initially expected to matter. The conversation also examines a larger problem in the AI industry: developers frequently select the largest available model before defining the problem they need to solve. Cloud services conceal the energy, infrastructure, privacy, and financial costs of those decisions. CrankGPT makes those costs visible—and makes a strong case for engineering systems around specific requirements instead of throwing more compute at every problem. We also cover Pinokio as an accessible entry point for experimenting with local AI, Whisper and whisper.cpp for speech recognition, modular voice-agent architecture, Raspberry Pi optimization, and why some supposedly ideal AI use cases are still better solved with a book, calculator, or conventional software. CrankGPT and build documentation:https://squeezlabs.github.io/handcrank/ CrankGPT:https://crankgpt.com/ Squeez Labs:https://squeezlabs.com/ Edge Voice Agent:https://github.com/ktomanek/edge_voice_agent Pinokio:https://pinokio.co/ whisper.cpp:https://github.com/ggml-org/whisper.cpp OpenAI Whisper:https://github.com/openai/whisper llama.cpp:https://github.com/ggml-org/llama.cpp Liquid AI:https://www.liquid.ai/ Gemma:https://deepmind.google/models/gemma/ DietPi:https://dietpi.com/ Raspberry Pi:https://www.raspberrypi.com/ Domesticating AI is a biweekly podcast about practical AI for developers. We help you brave the feral open-source AI landscape so you can tame it instead of getting dragged along by it. Keep your AI on a leash. Resources
You don’t need a DGX Spark to start running local AI. In this episode, Miriah, Matt, and Chris talk about realistic AI homelab hardware: old gaming PCs, refurbished desktops, Raspberry Pis, MacBooks, RTX cards, Mac Studio, DGX Spark, AMD Strix Halo machines, and cloud GPU rentals. The argument is simple: start with the hardware you already have, learn the constraints, then decide what to buy. We cover inference, small-model training, coding agents, local harnesses, OpenCode, Forge, LM Studio, Ollama, Docker, Podman AI Lab, vLLM, Lemonade, Unsloth, RunPod, Modal, and why hardware decisions should follow the workload instead of YouTube FOMO. We also get into the real reason local AI matters: data control, IP protection, regulated workflows, reproducibility, and understanding the engineering trade-offs instead of treating the model like magic. Comment question: What are you running local AI on right now, and what do you want to buy next? Links mentioned: LLMs in Production: https://www.manning.com/books/llms-in-production Andrew Ng Machine Learning Specialization: https://www.coursera.org/specializations/machine-learning-introduction fast.ai Practical Deep Learning for Coders: https://course.fast.ai/ NVIDIA DGX Spark: https://marketplace.nvidia.com/en-us/enterprise/personal-ai-supercomputers/dgx-spark/ NVIDIA Jetson Orin Nano: https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-orin/nano-super-developer-kit/ Framework Desktop: https://frame.work/desktop Beelink GTR9 Pro: https://www.bee-link.com/products/beelink-gtr9-pro-amd-ryzen-ai-max-395 RunPod: https://www.runpod.io/ Modal: https://modal.com/ OpenRouter free models: https://openrouter.ai/collections/free-models OpenCode: https://opencode.ai/ OpenCode Zen: https://opencode.ai/docs/zen/ Forge: https://github.com/antoinezambelli/forge LM Studio: https://lmstudio.ai/ Ollama: https://ollama.com/ Docker Model Runner: https://www.docker.com/products/model-runner/ Podman AI Lab: https://podman-desktop.io/docs/ai-lab vLLM: https://docs.vllm.ai/ AMD Lemonade Server: https://lemonade-server.ai/ Unsloth: https://unsloth.ai/docs Follow Domesticating AI: YouTube: https://www.youtube.com/@DomesticatingAI Apple Podcasts: https://podcasts.apple.com/us/podcast/domesticating-ai/id1873338950 Spotify: https://open.spotify.com/show/2WsAR4fvcXzp3vVZGVlkE2 Amazon Music: https://music.amazon.com/podcasts/193658ea-5d7e-40c5-b78c-dbc1f6ada65a/domesticating-ai Patreon: https://patreon.com/DomesticatingAIPodcast Keep your AI on a leash.