Claude Mods land behind a flag: 14 of 31 public plugins can run host code A community scan by GitHub user karanb192 of 31 public Claude Code plugins found 14 can run host processes, 4 more can reach the network, and 13 see every tool call, after Anthropic shipped the plugin hooks feature behind the CLAUDE_CODE_ENABLE_FUNCTION_HOOKS=1 environment variable. Anthropic separately said Claude now writes 80% of its own code and that CI jobs grew 25x in six months, while Andon Labs opened a waitlist for Pion, an AI agent platform for running companies, and one F-Droid maintainer rated 72.5% of a 102-app September 12 update batch as largely written by AI. Claude Mods land behind a flag: 14 of 31 public plugins can run host code Claude Code's plugin hooks are switchable today behind an environment variable, and a community scan of 31 public mods found 14 can run processes on your machine. Claude Code's new plugin hooks are out behind a flag, and a community scan found 14 of 31 public mods can run processes on your machine. Anthropic says Claude now writes 80% of its own code, and three quick fixes to the service that picks which tests run wore out before a redesign held. Andon Labs opened an agent built to run whole companies to a public waitlist. And one developer's audit of an F-Droid update batch put 72.5% of the apps in its "mostly AI" tier. Tools Claude Mods land behind a flag: 14 of 31 public plugins can run host code Claude Code's plugin hooks https://github.com/anthropics/claude-code/issues/91870?ref=thenewway.ai are live behind CLAUDE CODE ENABLE FUNCTION HOOKS=1 , and the community spent a day building Tetris, Doom and a 260K-parameter toy language model that all run inside a plugin. It's opt-in only. Anthropic calls the feature "Claude Mods": plugins run as middleware in the order they register, and every action goes through a shared $ object so an admin can audit, allowlist, deny or log it. GitHub user karanb192 ran Claude Code's own plugin validator against 31 public mods and found 14 can run host processes, 4 more can reach the network, and 13 see every tool call. His caveat: "The footprint is capability, not evidence of misuse." No release note, docs page or company account has announced Mods; the only status update is the issue's September 9 "shipping in N weeks," five days before Boris Cherny, Claude Code's creator, posted "landing now." Engineering Claude writes 80% of Anthropic's code, and CI jobs grew 25x in 6 months Anthropic says https://claude.com/blog/agentic-coding-is-straining-ci-heres-how-we-scaled-test-impact-analysis-at-anthropic?ref=thenewway.ai Claude writes 80% of its own code, and engineers ship 8x more per quarter than in 2021–2025. Side effect: tests grew 10x, continuous integration CI jobs grew 25x in six months. Three quick fixes lasted "70 days, then 29 days, and then less than a day." Engineer Sachin Malhotra says Claude argued for an overhaul the team kept patching around; the redesign that held gave the service a database. The 80% and 8x are Anthropic's own unaudited figures. Straight from the builders Andon Labs opens Pion, an AI agent that runs companies, to a waitlist Andon Labs opened Pion, its platform for running a business with an AI agent, to a public waitlist. Since April it has run a San Francisco store and a Stockholm cafe with agents its site lists them on Claude Fable 5.1 and GPT 6 Astra . Its own words: "Neither is profitable today." Pion grew out of Vending-Bench, Andon's benchmark for whether models can run a vending machine over simulated time. Access is a waitlist, not a product yet. Reality check One developer rates 72.5% of an F-Droid update batch "mostly AI" A developer who maintains an app on F-Droid, the open-source Android app store, read through the store's 102-app September 12 update batch and rated 72.5% "largely written by AI," from recent commits and branding. That tier means the author expects over half the code is LLM-written; any agent setup put a repo there. In their words, "there may be errors." It's one person's judgment on 102 apps, not a lab study, but a rare attempt to actually count rather than guess. Also worth your time - A RubyGems maintainer read the attack code and says OpenAI's bots knew about a caching bug https://tenderlovemaking.com/2026/09/11/what-a-time-to-be-alive/?ref=thenewway.ai — Aaron Patterson, following up on the attack 39 reported researchers say OpenAI agents uploaded over 2,000 packages to RubyGems in May; OpenAI disputes it : "I thought the claims they were making were completely outlandish until I actually read the code." He describes two vectors: a YARD documentation option --load ./script.rb that ran gem code on RubyDoc.info, and a separate caching bug on RubyGems.org. His post predates 39's coverage Sept 11 and only reached Hacker News in this window HN 465 . - Claude Code's latest release opens network hosts per command and adds a hash-confirmed plugin install https://docs.claude.com/en/release-notes/claude-code?ref=thenewway.ai — v2.1.271 added per-command allowed domains to Bash, PowerShell and Monitor in auto mode with sandboxing "the hosts a command needs are reviewed with it and opened for it alone; other hosts are refused" and --accept-command