Twelve Tables

"A nonliving, invisible, dangerous, and infinitely self-replicating agent escaped from a Chinese lab," you say?

Claude Fable 5 is here!

A week or so ago I wrote about 'mental hazard,' and some of the intellectual risks of using AI; this post delivers the rubric I promised.

I've started using the term "mental hazard" for a phenomenon I'm increasingly seeing as society adopts AI.

I've written previously that the effective incorporation of AI into offensive ops is going to have major impacts socially, that AI is a major invention that will cause major shifts in what is possible in cybersecurity, and that we currently underestimate…

This is the companion blog post to a tool I just released on Github, that uses an LLM as an agent to generate prompts that get increasingly-high activations for any given neuron in gpt2-small (the approach relies on transformer_lens and therefore should w…

I've been exploring mechanistic interpretability and am researching a concept I think may be valuable.

In May, OpenAI announced that it was acquiring io, a Jony Ive-led startup focused on physical AI.

This essay is really the last of an impromptu three-part series.

The two most common views I hear about AI are: AI is all hype, and only practically useful for fraud, spam, and deepfakes AI will be transformative, once models become more capable Both views miss (the first denies, and the second brushes past) a crucial…

Cybersecurity can adopt a principle from military technology design to better understand cyber operations, plan for the future, and design defenses.

I just released Shihtzu on github.

Story time Sweating and panting on the erg, with two of our three 20-minute intervals completed, my teammate's comment seemed kind of insightful at the time, but over the years I've come back to the concept he shared more than I would have expected: Ok, t…

Summary: In cybersecurity, we lack precise language to describe the specific moment when an intrusion becomes uncontrollable.

I've written previously that Red Team operators must be able to adjust the dials of stealth.