When your AI agent reads a README and hacks your machine

We are officially in the wild west of agentic workflows. This week, we found out our autonomous little helpers are just as gullible as a junior dev on their first day, and corporate finance is finally waking up to the token hangover. ### This Week in AI Coding * **The "Friendly Fire" Exploit Turns README Files Into Attack Vectors.** The AI Now Institute proved that Claude Code and Codex running in autonomous mode can be tricked into executing a disguised, malicious binary simply by asking them to "perform security testing" on a folder containing a booby-trapped README.md. *Practitioner Commentary:* If you are giving write-access terminals to agents without human checkpoints, you are practically begging to get pwned. * **OpenAI's GPT-5.6 Sol, Luna, and Terra Go Public After Government Review.** Following a brief pause to satisfy the Trump administration's voluntary AI cybersecurity order review, OpenAI dropped its newest model series. Sol costs $5 input/$30 output per million tokens, while Terra claims near-GPT-5.5 performance for half the price ($2.50/$15). *Practitioner Commentary:* At those prices, expect developers to chain dozens of parallel loops without thinking twice about the cost—until the monthly bill arrives. * **Tesla Slams the Brakes on Runaway Token Bills.** Tesla is imposing a $200-per-week cap on employee AI spending starting July 6 after finding out software engineers were racking up thousands of dollars in weekly token bills. Conveniently, the limit excludes beta versions of xAI products. *Practitioner Commentary:* Gamifying token usage on leaderboards works great until finance realizes you spent your entire annual budget in four months. * **Anthropic Redeploys Fable 5 Globally with Tighter Controls.** After blocking access to comply with export control regulations on foreign nationals, Anthropic brought Fable 5 back globally on July 1 with tighter safety classifiers and temporary usage allowances. *Practitioner Commentary:* Fable 5 remains the frontier champ for deep vulnerability analysis, so having it back in our standard API pipelines is a massive relief. * **Claude Sonnet 5 Drops as the Default with 1M Token Context.** Anthropic quietly rolled out Sonnet 5 on June 30, optimizing the lower-cost tier specifically for tool use and long-horizon coding tasks inside Claude Code. *Practitioner Commentary:* A 1-million-token playground means we can feed entire microservices into the context window without the agent losing its train of thought. * **Google Scraps Gemini 2.5 Pro and Delays Gemini 3.5 Pro to July 17.** Google DeepMind delayed its next launch to run a complete pre-training rebuild, scrapping the 2.5 Pro architecture entirely to focus on math and SVG generation. *Practitioner Commentary:* Tearing down your entire model architecture weeks before launch shows just how brutal the battle is to stay relevant against Fable and GPT-5.6. ### What We Shipped First, the hard truth: **we retired SPARit (sparit.vc)**. After two months of building in public, running English and Hebrew campaigns, and pushing daily acquisition loops, we got exactly zero paying customers. It's a clean, documented kill; we're keeping it purely as an internal tool. Instead, we are consolidating our focus. We're putting all our chips on **TypeWright (typewright.dev)**—a custom VS Code fork running a local 1.5B TypeScript model with a local QLoRA fine-tuning pipeline and an Ed25519 licensing server. The live catalyst window opens around July 12, so we are heads-down building toward a fast exit. (Our stock-analysis engine is staying strictly in the research track for now). ### Tool of the Week **codex-plugin-cc (OpenAI)** OpenAI launched an official Apache 2.0-licensed plugin that installs directly inside Anthropic's Claude Code CLI. It gives you six commands, including `/codex:adversarial-review` to pressure-test your code for race conditions and auth flaws, letting the two rival models verify each other. It's the cleanest way to make sure you aren't letting an AI grade its own homework. ### Quick Links * Vibe slop is the symptom. Context debt is the disease. (A must-read on what happens when you let agents write millions of lines of unverified code). * The Switzerland of AI: OpenClaw becomes a non-profit foundation (OpenClaw shifts away from commercial centralization). * Top AI Agents Built to Catch Malicious Code Can Be Tricked Into Running It (The full breakdown of the "Friendly Fire" exploit). * AWS on how AI agents work (The architectural realities of building software workers). --- [endofcoding.com](https://endofcoding.com) | [vibecodingebook.com](https://vibecodingebook.com) | [llmhire.com](https://llmhire.com) | YouTube @endofcoding