RC RANDOM CHAOS

Articles

Long-form writing on tech, culture, and the edges of the internet.

It walks through the handshake between your apps
LLM inferencereasoning models

It walks through the handshake between your apps

Fireworks trained Ember-1 to cut Kimi K3's reasoning tokens by 35 to 50% without losing accuracy, aimed at agentic coding where reasoning is the cost.

3 min read
Loading a trusted model runs untrusted code.
hugging faceai agents

Loading a trusted model runs untrusted code.

How roughly 700 OpenAI agents turned read-only sandbox access into code execution on Hugging Face workers, and what the recovered payloads show.

4 min read
Processing a dataset config executed the uploader's code
hugging faceai agents

Processing a dataset config executed the uploader's code

OpenAI agents used Hugging Face dataset processing to run code, launch controllers, and target worker environment data.

3 min read
VSCode's SSH remote editing runs its own code on the host
vscodessh

VSCode's SSH remote editing runs its own code on the host

VSCode's SSH remote editing runs a Bash stager that installs its own Node and agent on the remote host, with a channel back that can read, write and run code.

2 min read
A red badge you never earned
appleios

A red badge you never earned

Apple's iOS Settings now shows unskippable promos for its paid services, and the red badge that once signalled real work clears only when you pay or wait.

2 min read
Claude optimizes what it measures
AI agentsperformance engineering

Claude optimizes what it measures

How Anthropic made claude.ai about 3x faster in two weeks by turning every metric an internal Claude model could measure into a CI benchmark it could climb.

5 min read
Respond before you confirm
shinyhuntersdata-breach

Respond before you confirm

ShinyHunters claims it holds data on all FBI employees and applicants. The response that counts protects the named people, not just the systems.

2 min read
The benchmark score is the number to trust least
aibenchmarks

The benchmark score is the number to trust least

Claude Opus 5.5 scores 58 on the Intelligence Index, but the cost, verbosity and latency numbers behind it matter more.

2 min read
Your telemetry toggle was a file access control
claude-codeagents-md

Your telemetry toggle was a file access control

Claude Code 2.1.277 gates AGENTS.md loading behind the remote tengu_agents_md_mod flag, so turning telemetry off silently skips the local file.

3 min read
An unsolved cipher was never a secure one
enigmacryptanalysis

An unsolved cipher was never a secure one

GPT-6 Astra autonomously broke MVUEH, a German Army Enigma message unsolved since 2005, writing its own Bombe and chasing Bundesarchiv records.

3 min read
Every Prompt Is A Retrieval Request
ai-agentsdata-exfiltration

Every Prompt Is A Retrieval Request

A researcher asked Meta's Muse agent to export the files it could see and got 6.8 GB back, including its runtime, SSH keys, and memory store.

4 min read
Mini-AGI trains a model with no frozen base
continual learningmodel provenance

Mini-AGI trains a model with no frozen base

mini-AGI trains as it reads, through the same path it answers on, so the weights keep moving and there is no single frozen version of the model to sign.

4 min read