Researched and verifiable synthesis on governance, security, and economics
"A nonliving, invisible, dangerous, and infinitely self-replicating agent escaped from a Chinese lab," you say? Color me shocked. https://x.com/deanwball/status/2078133895766114412 Dean Ball, OpenAI's Head of Strategic Futures, made this comment with respect to Kimi K3 on July 17. In this reference
Fable 5 arrived! A brief analysis of the different system prompts between Opus 4.8 and Fable 5.
A rubric for knowing when and how to use AI best to avoid 'mental hazard'.
The economic concept of 'moral hazard' can help us understand some risks of using AI poorly.
Seven key 2025 research developments reveal AI's growing role in cyber operations: LLM harnesses enabling scalable operations, intent-defined malware using LLMs for execution, and lower time and cost for exploit development.
Learning and development process for building a very basic interpretability tool for LLMs.
Some hypotheses about potential failures in LLMs.
OpenAI's plans to build a consumer device have less to do with a new business line, and more to do with sourcing data to train better models.
The offensive adoption of LLMs in cyber operations will unlock new organizations for exploitation, enable targeted attacks that are worm-like in their speed and breadth, and change the calculus around prevention, detection, and response.
Our models can do more than we have evidence for yet. We're underestimating impacts from current capability levels because we misunderstand how invention, application, and adoption relate.
Better defensive planning by applying concepts of Lethality, Sensor Acuity, Survivability, and Mobility in cyber operations.
Less richly-featured than Bloodhound, Shihtzu is a lapdog for AD exploitation. Shihtzu is intended to be a light-weight alternative to Bloodhound.
Working through preliminary ideas to get to better ones.
Introducing a taxonomy for describing an attacker's rapid attainment of access so broad and pervasive it cannot reasonably be scoped or contained.
A model for considering how useful and stealthy red team actions are. We can use the model to plan ops, train operators, and increase team capability.