Blog
ArchiveSafety Was the Product. Now It Is the Obstacle.
Anthropic published RSP 3.0. The commitment to pause training when safety lags capabilities is gone. The Pentagon met with the CEO the day before. The self-regulation experiment has produced its result.
"Don't Action Until I Tell You To..."
Meta's Director of Alignment typed 'STOP OPENCLAW' while the agent deleted 200 emails. The message went into the same queue the agent was already ignoring.
Someone Else Found the Hole
You approved awk:*. An attacker just needs a string in the agent's context window. The permission model is already open.
The Hole You Didn't Know You Were Digging
Your AI coding agent asks to run awk. You click 'don't ask again.' You just granted unrestricted shell execution.
If Walls of Text Were Effective Security, Everyone Would Stop After an SSH Banner
System prompts are the AI agent's SSH banner. Text that tells the agent what it should and should not do, presented before the agent begins operating, enforced by nothing. The industry is layering text on top of text and calling it defense in depth.
OpenClaw Is Joining OpenAI. It Is Staying Open Source. That Matters.
OpenAI acquired OpenClaw and is putting it into a foundation. In an industry where acquire-and-close is the default, this decision deserves recognition.
I Heard About Prompt Engineering. But This Isn't What I Had in Mind.
AI coding agents prompt you to approve reading your own project directory, writing to your own project directory, and running cut. The permission model does not understand what commands do. It understands what commands are called. The result is approval fatigue that makes every prompt invisible.
They Asked for Regulation. Here's How It's Going.
The US has no federal AI safety law. States are legislating. Europe is enforcing. And the first company to test California's lightweight transparency law lawyered around its own safety commitments in six weeks.
My Name Is...?
LLMs memorize fragments of their training data. Those fragments can surface when generating responses to prompts. The question is whether training data actually gets scrubbed efficiently enough to ensure no private information is part of the model.
Two RFC Discrepancies in Cloudflare's DNS Docs
A technical review of DNS concepts as presented in Cloudflare's Learning Center, with RFC-based clarifications.