RC RANDOM CHAOS
RC RANDOM CHAOS

Tech · Culture · Fiction

Clef shipped, and attackers rented a decision model Article

Clef shipped, and attackers rented a decision model

Open-weight decision models plus cheap RL fine-tuning let anyone point a goal-seeking AI agent at any reward - including offensive ones. What that means for security.

The lie and the order wear the same font
Article prompt injection

The lie and the order wear the same font

Gemini 4 Argon merges instruction and data in one channel, so controls inside the model cannot enforce a boundary that no longer exists.

Article macos-security

Golden Gate's crashes cluster in exploit-grade code

macOS Golden Gate's reported instability sits in the same subsystems - WebKit, ImageIO, IOKit, XNU - that carry every documented macOS exploitation chain.

MicroLLM Lab runs seven models without touching the network
Article browser LLMs

MicroLLM Lab runs seven models without touching the network

In-browser tiny LLMs keep prompts off the network, but same-origin scripts, local storage, WebGPU fingerprinting, and weak model safety reshape the risks.

Your best lock was a price tag
Article social engineering

Your best lock was a price tag

GPT 6.1 Sol cuts near-frontier AI cost fivefold, collapsing the economic scarcity that quietly propped up social engineering and phishing defenses.

Anthropic, OpenAI, and DeepMind grade their own models' danger
Article AI safety

Anthropic, OpenAI, and DeepMind grade their own models' danger

AI labs publishing dangerous-capability evals turns safety disclosure into marketing, and the self-graded scorecards drift toward a race to the bottom.

Article AI governance

Unsealed court briefs turned training data into evidence

Unsealed briefs in the authors' case against Microsoft and OpenAI reveal training-data provenance as a board-level liability surface, not a technical detail.

The Wire — latest

All →

Stay in the loop

New writing delivered when it's ready. No schedule, no spam.