For a compact date-based index, see Archives. Browse by topic in Tags.
Running an open model on your own server doesn't tell you what's inside the weights. Finding and fixing model backdoors is following the path software security already took, but models are harder both to read and to fix.
There's a lot of talk these days about copying commercial software with AI, but what AI reads out of a binary isn't the original. It's code with the blanks filled in plausibly. That may be enough for copying, but when you're looking for vulnerabilities, those blanks become the problem.
AI is producing vulnerabilities and exploits faster than ever, but patches aren't reaching users any faster to match. I took the patch gap and blast radius questions I raised in March and checked them against seven months of new numbers.
AI has made finding bugs easy. Filtering them, fixing them, and handing them to maintainers still takes a lot of work.
Seeing ads under ChatGPT answers got me thinking: can AI change how people think without a single lie, just by choosing what to include, what to leave out, and what to show first? And what about agents? Here's what the research and regulation say so far.
The code and docs my agents produced were neatly organized, but that didn't mean I understood them. Starting from a line Karpathy quoted and the idea of cognitive debt, I look at how to bring what agents produce in line with what people actually understand.
However carefully you write down what you want, it's hard for the person you hand the work to to follow your thinking 100%. Agents are no different. Here's what needs to be aligned up front, who should review the code an agent writes, how much to automate, and who owns the judgment that's still left after you automate.
Somewhere around the third time I paste the same prompt, I want to make it a skill. When to do that, what goes inside a skill, how to evaluate skills stacked like blocks and the seams between them, and which model to grade with.
A session can burn through hundreds of thousands of tokens and the model usually comes out unchanged. So I thought about what to pick out of a session and keep, and where to build it up: prompt, context, harness, or loop.
The time from vulnerability disclosure to exploit has shrunk from 2.3 years to 1.6 days. AI is widening the gap between attack and defense.
Google's AP2, Cloudflare's NET Dollar, ChatGPT Instant Checkout: a roundup of the agent payment announcements that poured out within two weeks in September, with my notes on why agents end up spending money and what worries me about it.
Team Atlanta wins 1st place in DARPA's AI Cyber Challenge (AIxCC), demonstrating AI-driven vulnerability detection and patching.
If an AI learns from the same data as me, how do I prove which one is the real 'me'? Random thoughts on fake information and phishing getting more convincing with generative AI, why AI red teams need to look at safety too, and a study that jailbroke LLMs with low-resource languages.
A friend's company got hit by eCh0raix ransomware. Here's what I learned about ransomware recovery, negotiation models, and community debates.
