RC RANDOM CHAOS

Articles

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

Your watermark is a spymark
watermarkingprivacy

Your watermark is a spymark

Watermarks that embed a hidden per-user identifier become spymarks: what the payload size tells you, and why metadata stripping does not remove them.

3 min read
MCP is an attack surface, not a feature
MCPAI agents

MCP is an attack surface, not a feature

An engineer argues the Model Context Protocol is obsolete now that capable agents can call HTTP APIs and CLIs directly, and most MCP servers can go.

3 min read
Rewriting weights live on 8GB of VRAM
machine-learningcontinual-learning

Rewriting weights live on 8GB of VRAM

mini-AGI trains a byte-level MoE model from scratch on an 8GB GPU, keeping weights on disk and paging experts in, learning as it reads.

5 min read
ZuckOff detects the camera outside your access model
privacybluetooth

ZuckOff detects the camera outside your access model

ZuckOff flags Ray-Ban Meta, Oakley Meta and Snap Spectacles by their Bluetooth signature, a signature-based detector for the recording devices your access model never sees.

2 min read
ChatGPT already knows your other browsing.
ChatGPTprivacy

ChatGPT already knows your other browsing.

OpenAI's __obi cookie ties your ChatGPT account to browsing on advertiser sites, and it fires even if you grant analytics but refuse marketing.

4 min read
Most automation doesn't need a smart model
on-device AItool calling

Most automation doesn't need a smart model

Cactus Needle 3 is an 8-29 MB offline model for tool calls, extraction and embeddings, with triggers and confidence gating you have to wire carefully.

3 min read
One write past the chunk, read-write on the repo
securityexploitation

One write past the chunk, read-write on the repo

How Hacktron chained a libheif heap overflow and an OpenAI SSO misconfiguration to reach OpenAI's internal repo in under 72 hours, with AI doing the exploit work.

4 min read
Microsoft called it theft; the crawler maps your attack surface
AI scrapingcopyright

Microsoft called it theft; the crawler maps your attack surface

Unsealed NYT v. OpenAI/Microsoft filings show a 93% click-through drop, internal "theft" admissions, and the scraping paths that already hold your content.

3 min read
OpenJev benchmarks against a value it never explains
local-llmlogits

OpenJev benchmarks against a value it never explains

OpenJev times two ways of reading a local model's option probabilities in-browser, but never explains the Jev value its accuracy table scores against.

2 min read
Parameter count is the wrong number to watch
model compressionlocal inference

Parameter count is the wrong number to watch

PrismML's Ternary Bonsai 2 27B fits a 27B multimodal model into 5.9GB, and its footprint and energy numbers matter more than parameter count.

2 min read
Passkeys don't stop account takeover
passkeysauthentication

Passkeys don't stop account takeover

Passkeys stop login-flow phishing but leave SMS and email recovery intact, so account takeover and lockout stay the real risks for most users.

2 min read
The hallucination was not the failure.
military AIAI hallucination

The hallucination was not the failure.

An AI-generated report nearly sent US forces to board a Chinese ship. The failure was how the hallucination was packaged as trusted intelligence.

3 min read