Research Notes

The lab's working notebook: field notes on AI security, kept in the open.

• What People Built With Fable 5
• What Fable 5 Holds for Us
• The Same Query, Beneficial or Dangerous
• Export Control as an AI-Safety Lever
• The Rebuttal: The Capability Was Already Diffuse
• The Word Security Appears Zero Times
• Equation 1 Is the Root Cause of Prompt Injection
• The Amazon Classifier Bypass
• Project Glasswing and Gated Frontier Access
• Two Doors on One Model: The Fable and Mythos Split
• Mythos-Class: A Tier Above Opus
• Package Hallucination and Slopsquatting
• Zero-Click Agent Data Exfiltration in Connected Assistants
• IMO Gold and the Stargate Buildout
• Gemini 3 Pro and the Agentic Default
• Sora 2 and Synthetic Reality
• Claude 4 to Opus 4.5, Agents and Alignment Under Load
• GPT-5 and the Unified Router
• gpt-oss and OpenAI's Return to Open Weights
• Kimi K2 and Open Agentic Intelligence
• Llama 4 Scout and Maverick, Open MoE and Long Context
• Grok 4 and Native Tool-Using Reasoning
• Grok 3 and Compute as Strategy
• Gemini 2.5 Pro and Thinking at Long Context
• Claude 3.7 Sonnet and the Rise of Claude Code
• GPT-4.5 and the Limits of Pure Scale
• Deep Research and Autonomous Multi-Step Investigation
• Operator and the Browser as an Action Surface
• DeepSeek-R1 and the Opening of Reasoning
• Morris II, a Research Prompt-Injection Worm
• Training-Data Poisoning and Backdoors
• Malicious Model Uploads on Model Hubs
• PoisonGPT and the Tampered-Model Demonstration
• Data-Exfiltration Risk in Slack AI Connected to Private Data
• Data Exfiltration via Rendered Markdown Images
• Indirect Prompt Injection via Web Pages and Email
• Air Canada Held Liable for Its Chatbot's Answer
• The EU AI Act Enters Into Force
• o3 and the ARC-AGI Breakthrough
• The 2024 Nobel Prizes and AI's Scientific Arrival
• AlphaGeometry and Olympiad Geometry Reasoning
• AlphaFold 3 and Modeling the Molecules of Life
• Model Context Protocol and Standardized Tool Access
• Anthropic Computer Use and Agentic GUI Control
• DeepSeek V2 and V3 and Efficient Scaling
• Llama 3 and the 405B Open Frontier
• Mixtral 8x7B and Open Sparse Mixture of Experts
• Devin and the AI Software Engineer Claim
• OpenAI o1 and Hidden Chain-of-Thought
• GPT-4o and Native Omni-Modal Interaction
• Claude 3 and the Opus, Sonnet, Haiku Tiering
• Sora and the Arrival of Usable Text-to-Video
• Gemini 1.5 and the Million-Token Context Window
• DAN and Roleplay Jailbreaks
• The Chevrolet Dealership Chatbot and the One-Dollar Car
• Early Prompt-Injection Demonstrations
• Samsung Source Code Leaked Into ChatGPT
• The ChatGPT History Leak From a Redis Bug
• Bing Chat, Sydney, and Early Alignment Failures
• Code Llama for Programming
• QLoRA and Cheap Fine-Tuning
• Mistral 7B and Efficient Open Models
• Llama 2, Deliberately Open
• ReAct: Reasoning and Acting
• Toolformer and Self-Taught Tool Use
• LangChain and the Agent Tooling Boom
• BabyAGI and Task-Loop Agents
• AutoGPT and Autonomous Agents
• Google Bard Enters the Race
• Anthropic Ships Claude
• The GPT-4 System Card and Red-Teaming as Norm
• GPT-4 Arrives
• Vicuna and the Chat-Quality Race
• Alpaca and Cheap Instruction Tuning
• LLaMA and the Open-Weights Leak
• Galactica and Its Withdrawal
• The Emergent Abilities Debate
• The ChatGPT Launch
• Whisper and Open Speech Recognition
• Midjourney and Product-Led Image Generation
• Stable Diffusion and Open Image Generation
• Parti and Autoregressive Image Generation
• Imagen and Photorealistic Text-to-Image
• Flamingo and Visual Language Models
• Minerva and Quantitative Reasoning
• Self-Consistency Decoding
• Chain-of-Thought Prompting
• RLHF as an Alignment Method
• InstructGPT and Instruction Following
• PaLM and the 540B Frontier
• Chinchilla and Compute-Optimal Scaling
• AlphaFold 2
• Megatron-Turing NLG 530B
• Jurassic-1 (AI21)
• Gopher (DeepMind)
• T0 and Multitask Prompting
• FLAN and Instruction Tuning
• Retrieval-Augmented Generation (RAG)
• Switch Transformer and Sparse MoE
• GitHub Copilot
• OpenAI Codex
• DALL-E
• CLIP and Image-Text Alignment
• The Rise of Prompt Engineering
• In-Context Learning as a Phenomenon
• Scaling Laws for Neural Language Models
• GPT-3 and Few-Shot Learning
• GLUE and SuperGLUE Benchmarks
• Megatron-LM and Model Parallelism
• XLNet
• T5 and Text-to-Text
• GPT-2
• GPT-1
• DistilBERT and Distillation
• ALBERT
• RoBERTa
• BERT
• ULMFiT and Transfer Learning for NLP
• ELMo and Contextual Embeddings
• word2vec and GloVe Embeddings
• Sequence-to-Sequence Learning
• Bahdanau Attention as the Precursor
• The Transformer Architecture, One Year On