Publications

Open research on the security of autonomous AI systems, released in measure.

• Fable 5 in the Wild: What a Mythos-Class Model Did in Its First Month
• Released in Measure, by Force: The Fable 5 and Mythos 5 Suspension
• The Unregulated Machine: A Security Reading of Attention Is All You Need
• Autonomy as Attack Surface: Security in the Agentic Era
• The AI Supply Chain: Model Weights, Malicious Uploads, and Slopsquatting
• Memory Poisoning in Agentic Systems
• Red-Teaming and Evaluations for Language Models
• Data Exfiltration via Markdown Image Rendering in Chat UIs
• Indirect and Second-Order Prompt Injection via Retrieved Content
• The Model Context Protocol and the Agent Tool Layer
• Reasoning Models and Test-Time Compute: o1, o3, and DeepSeek-R1
• Multimodality: From CLIP to GPT-4o and the New Input Channels
• Jailbreaking LLMs: Methods and Why Alignment Training Does Not Close the Gap
• Prompt Injection: A Systematic Taxonomy
• Tool-Using Agents: Toolformer, ReAct, and the Action Boundary
• Direct Preference Optimization: Alignment Without a Reward Model
• Mixture-of-Experts at Scale: What Mixtral and DeepSeek Changed
• LoRA and the Economics of Parameter-Efficient Tuning
• Retrieval-Augmented Generation and the Injection Surface It Opened
• Constitutional AI: Principles Over Labels
• Chain-of-Thought Prompting and the Roots of Reasoning
• InstructGPT and RLHF: Aligning Models to Instructions
• Chinchilla and Compute-Optimal Training
• Neural Scaling Laws and the Kaplan Predictions
• GPT-3 and In-Context Learning
• GPT-2 and the Staged-Release Safety Debate
• BERT and the Bidirectional Pretraining Turn