OpenAI launched always-on agents — then fired three of its own safety researchers

I ship code with AI agents all day, and this week both big labs reminded me why I don't fully trust any of them unsupervised. OpenAI handed out always-on agents with their own cloud computers — then fired three of its own safety researchers for talking to the outside watchdogs who'd normally catch something going wrong. Meanwhile Claude Code and Cursor both shipped tools aimed at the exact same anxiety: something is going to slip past you, so here's a bot to watch for it. ## This Week in AI Coding - **OpenAI launched "dots" — always-on ChatGPT agents with their own cloud computer and browser.** Announced at DevDay on Sept 29, dots run on GPT-6 Astra, connect to 4,000+ apps, and work toward a standing goal between your prompts instead of waiting for the next one. Rolling out to ChatGPT Pro and Business Premium (Pro excludes the EEA, Switzerland, and the UK for now). If you've been waiting for "set it and forget it" agents to stop being a slide at a keynote, this is the first mainstream one that isn't. - **OpenAI's Agents API went public beta the same day** — a managed version of the Codex harness with computer use and multi-agent support built in, so you get OpenAI's session orchestration and context-compaction instead of rolling your own. Worth a look if you've been hand-rolling agent loops; it's the first time they're selling the plumbing, not just the model. - **OpenAI fired three safety researchers — Jasmine Wang, Tomek Korbak, and Mikita Balesni — for sharing confidential information with outside AI-safety groups** (one of them handled OpenAI's liaison work with METR and Redwood Research, the orgs that do a lot of the independent model evals you've probably seen cited). First reported by the WSJ Oct 1-2. The optics are rough: a company rolling out always-on agents this same week just fired the people whose job was partly to keep outside eyes on it. - **Claude Code shipped Mods (v2.1.287, Oct 1)** — plugins can now reach deeper into the CLI/desktop app: new UI, rewritten prompts, replaced built-ins. The first built-in mod is "You should know" — a side agent that watches the main session and flags things you or Claude might be missing. `/plugin enable cc-plugin-you-should-know@builtin` if you want it on. - **Cursor shipped Rollouts and Security Review (Sept 23)**, two bots for Teams/Enterprise plans. Rollouts watches a PR's change through deploy and reports per-environment health (verified / regression / inconclusive) against your logs and metrics. Security Review posts one comment per PR naming exploitable bugs — injection, auth bypass, secrets, SSRF — with a severity and a proposed fix. Direct answer to "my agent shipped it, who's checking it." ## What We Shipped - Closed a month-old mystery on the LLMHire traffic spike from August: it was a headless crawler out of a residential-proxy pool, not real visitors — real human traffic has been a steady 15-45/week the whole time. - Finally got Veridux Analytics installed on Metavchim, closing a tracking gap that's been open since launch. - Honest miss: the automated daily billing tracker failed its run again on Sept 27, and this week's site-health sweep came back 33/36 — EndOfCoding's landing page still has leftover placeholder copy ("coming soon") I need to clean out. ## Tool of the Week **Claude Code's "You should know" mod.** It's a second agent riding along in the same session, watching for the stuff you or the main agent might miss — a quiet second pair of eyes rather than another chat window to babysit. If you're running Claude Code solo without a human reviewer most days (I am), it's worth the one command to turn on. ## Quick Links - [OpenAI DevDay 2026: dots, Agents API, and the rest of the keynote (CNBC)](https://www.cnbc.com/2026/09/29/openai-devday-2026-live-updates.html) - [OpenAI fires three safety researchers over alleged leak (Yahoo Tech)](https://tech.yahoo.com/ai/articles/openai-fires-three-safety-researchers-181204664.html) - [Cursor: Rollouts and Security Review, full writeup](https://cursor.com/changelog/rollouts-and-security-reviewer) - [Claude Code changelog — v2.1.287, Mods + You should know](https://code.claude.com/docs/en/changelog#2-1-287) --- Building a portfolio of AI products solo, in public: [endofcoding.com](https://endofcoding.com) · [vibecodingebook.com](https://vibecodingebook.com) · [llmhire.com](https://llmhire.com) · [@endofcoding on YouTube](https://www.youtube.com/@endofcoding)