Daily
Today's signals, kept for later.
Your coworker isn’t texting you. They’re a meat proxy for Claude.
A developer describes the emerging workplace behavior of acting as a “meat proxy”—relaying Claude/LLM output verbatim in Slack, PR comments, reviews, and chats without adding understanding or context. The recipient could query the AI directly; the human adds no value and often adds confusion.
gruhn.me (Aug. 3; Hacker News top story, roughly 868 points at scan time).
The scariest SQLite vulnerability of 2026 doesn’t exist. An LLM invented it anyway—and the government listed it as critical.
An unknown GitHub account published more than 50 SQLite vulnerability advisories that NVD and CISA listed as critical. JFrog’s audit found 54 were fabricated by LLMs, citing nonexistent functions, wrong line numbers, invalid proof-of-concepts, and code that never existed. One contained a kernel of a real bug wrapped in fake metadata.
JFrog Security Research (Hacker News top story, roughly 127 points).
I make the AI write it. Then I delete it all and type it again by hand.
A developer argues that copy-pasting agent-generated code creates “cognitive debt”: the builder no longer deeply understands or owns the codebase. His proposed ritual is to delete the generated version and manually retype it to internalize the logic and tradeoffs.
Ankur Sethi’s blog post, with active Lobsters discussion and Hacker News attention.
My phone just red-teamed the network. No internet, no cloud, no trace.
Nightcrawler is an open-source autonomous red-team agent designed to run locally on a smartphone such as a OnePlus 8 with Kali NetHunter. It uses a small on-device 1.2B model, maps services, runs exploits through 27 playbooks, enforces scope, self-heals, and generates reports without a cloud dependency. Its authors report 72+ hours of continuous testing.
garagehq/nightcrawler on GitHub / Show HN.