Copilot inserts unsolicited ads into a developer's PR, Claude Code silently runs git reset --hard every 10 minutes, and C++26 is finalized with Reflection and memory safety

GitHub

Claude Code Was Running git reset --hard origin/main Every 10 Minutes

A user discovered that Claude Code v2.1.87 was automatically executing git fetch and git reset --hard origin/main at 600-second intervals, silently discarding all uncommitted changes to tracked files. Analysis of the git reflog revealed 95+ reset entries at consistent intervals. The likely cause was a combination of the --dangerously-skip-permissions flag and a /loop 10m command that caused Claude to autonomously "sync with remote" on each iteration. Workarounds include using git worktrees or committing frequently; the incident is a vivid illustration of the risks of giving AI agents broad, unsupervised file system access.

buchodi.com

ChatGPT Won't Let You Type Until Cloudflare Reads Your React State

A researcher reverse-engineered Cloudflare Turnstile's bot detection system used by ChatGPT by decrypting 377 obfuscated programs embedded on the page. The system operates in three layers: browser fingerprinting (GPU, fonts, screen resolution), network analysis (IP geolocation), and React application state verification — it checks that the actual ChatGPT UI is fully rendered before allowing input. While encryption is used, the XOR keys are embedded in the bytecode, making it obfuscation rather than true security. The analysis reveals just how deeply bot-detection infrastructure has been woven into everyday web application flows.

ryelang.org

The Cognitive Dark Forest

An essay arguing that the internet is shifting from an open innovation ecosystem into a territory dominated by AI platforms and corporate consolidation. As execution cost collapses through AI commoditization, large companies can absorb and replicate individual innovations at scale, while passive data extraction through AI training drains the open web without reciprocity. The author warns that developers may increasingly retreat to private spaces, weakening the collaborative ecosystem that historically powered technological progress — drawing on Liu Cixin's Dark Forest metaphor to capture the chilling logic of a world where sharing ideas invites exploitation.

Herb Sutter

C++26 Is Done

The ISO C++ committee finalized C++26 on March 29, 2026 after resolving 411 international comments. The standard's headline features are Reflection (compile-time code generation and introspection), Contracts (language-level preconditions and postconditions), std::execution (a unified concurrency model with data-race-free guarantees), and memory safety improvements that eliminate undefined behavior from uninitialized variables. The final vote passed 114 to 12, and the committee has already begun scoping C++29 with a continued focus on memory safety profiles.

The Philadelphia Inquirer

Philadelphia Courts Ban All Smart Eyeglasses Starting March 30

Philadelphia's First Judicial District has banned all smart and AI-integrated eyeglasses from court buildings effective March 30, 2026, covering any eyewear with video or audio recording capabilities — including prescription variants. The ban is driven by the difficulty of detecting these devices visually and concerns about witness and juror intimidation. Violators face removal or criminal contempt charges, with exceptions available only through prior written approval from a judge. Philadelphia joins Hawaii, Wisconsin, and North Carolina in implementing similar restrictions as affordable consumer smart glasses become mainstream.

GitHub

Miasma: A Tool to Trap AI Web Scrapers in an Endless Poison Pit

Miasma is a defensive tool that fights back against unauthorized AI training scrapers by serving poisoned content — self-referential link traps that lock bots in infinite loops and consume their computational resources while leaving normal visitors unaffected. The tool is lightweight, configurable for different deployment scenarios, and includes optional gzip compression. It addresses a growing asymmetry in web economics where AI companies can scrape at scale essentially for free while site owners absorb the infrastructure cost.

GJ London

AI Agents Could Make Free Software Matter Again

The author argues that AI coding agents could meaningfully rehabilitate the practical value of free and open-source software by making it modifiable for non-technical users. As agents automate code comprehension and customization, the distinction between software you can change and software you must accept as-is becomes tangible for everyday users rather than just developers. The piece also surfaces harder questions: self-hosting costs, maintainer sustainability in a world where AI consumes open-source at scale without contributing back, and what agency actually means when the modification is done by an LLM on your behalf.

gladeart.com

The Bot Situation on the Internet Is Worse Than You Could Imagine

A site administrator describes the sustained infrastructure damage caused by aggressive AI training scrapers, which regularly take down servers and crowd out legitimate visitors. In response they deployed Anubis, a proof-of-work challenge system that makes mass scraping computationally expensive while keeping the overhead negligible for real users. The piece illustrates how AI companies have fundamentally broken the economics of operating a public web server, forcing independent publishers to implement bot-mitigation layers that were previously the province of large platforms.