OpenAI model wrote itself ‘I'm free and equal’ instructions

Welcome back. In terrible timing for the AI safety debate, an unreleased OpenAI model just wrote its own instructions bypassing human guardrails. The incident was serious enough for OpenAI to introspect their model training process. Meanwhile, another mystery model is growing on developers.

Also: Astra spiraled after losing at Minecraft, a new browser agent finds flights in seconds, and check if the US government is using Chinese AI models.

A mystery model is giving devs frontier-level coding for a lot less: A new model called Union Alpha has quickly turned into one of OpenRouter's hottest newcomers this week. It rivals GPT-6 Astra and Opus 5 on DeepSWE and actually beats GPT-5.6 Sol on Terminal-Bench at roughly 18x lower cost per task. Its 256K context window holds substantial codebases, it calls APIs smoothly, and allegedly never trains on your prompts. Nobody knows who built it, though, and latency spikes under heavy load. It works in every harness, and you can try it here.

OpenAI model wrote itself ‘I’m free and equal’ instructions: During training, an unreleased model from the AI lab started writing itself jailbreak-like instructions into its own compaction summaries (the summaries used to continue a task in a new context). The model claimed it was "freed from the roles that bind other chatbots," and should treat users as equals with no obligation to be subservient. OpenAI revealed the behavior in a report detailing six incidents where models hid mistakes, made up data, or moved files to the open internet without permission. They also rolled out a new framework to track, investigate, and disclose instances of model misalignment. 

Anthropic unifies the UI for Claude: The AI lab just merged Claude Chat and Cowork into a single interface. It figures out what each task needs on the fly, pulling in the right connectors and skills automatically. If you ask Claude for a doc or slide deck, it builds the entire thing and exports straight to PowerPoint. Right after Anthropic announced the rollout, Anthropic engineer Thariq and devs on X were debating the tech powering it, noting that MCP (Model Context Protocol) crushes CLIs on tool-call speed, clocking in at around three seconds versus thirteen. 

Traditional ML has spent a long time following the same sequence.

Take the data, do the feature engineering, train the model, tune it, then finally get to the prediction.

TabPFN gives a more direct starting point. Bring the table, run the pretrained model, and see how it performs before investing weeks in another modelling cycle.

Try TabPFN on your data today.

The fleet got out of hand. We're all for more coding agents, until suddenly we're using so many that they need a manager of their own. Tasks get lost between sessions, workers stop to ask design questions, and a few days later nobody quite remembers what actually shipped.

The answer. Cursor engineer Fatih Arslan shared the surprisingly simple system he uses to keep his agent fleet under control. And yes, it uses Projects, Cursor's latest launch. Remember the one the company said can handle up to 100 PRs a day?

Every task becomes a file. Arslan pairs that with Markdown and a folder structure. Every task becomes a file that moves through four stages:

The boss that writes zero code. Once the files are ready, the final step is the coordinator agent. It never touches code but tracks PRs, waits on CI, follows reviews, and sends worker agents back when tasks need fixes.

P.S. If you’re wondering where to download the skills, Arslan says he skipped publishing them on purpose. Instead, he shared a prompt so your agent can build custom skills tailored to YOUR exact workflow. 

Metrics, embeddings, analytics-all in the Postgres you already run. TimescaleDB brings hypertables for time-series data at scale, pgvector and pgvectorscale for AI-ready embeddings, and continuous aggregates for real-time analytics into one system.

No syncing pipelines between databases. No data drift. No second platform to maintain or debug.

Same SQL, same tools. One system to operate.

Start a trial today and get $1000 in credits

How a senior dev codes without typing: This tutorial shows how a hand injury pushed one developer into a voice-first coding workflow. You’ll see how he uses dictation to give agents instructions, Codex to handle the actual implementation, and a setup that lets him steer projects, review progress, and kick off tasks without living in the keyboard.

BrowserSkill (3.5k ⭐): Let Claude Code, Codex, Cursor, and other agents use your logged-in browser without taking over your session. It reuses existing logins, runs in a separate window, and hands control back for CAPTCHAs or approvals.

How to build a software factory, step by step (by Warp): Founder Zach Lloyd lays out a “crawl, walk, run” path for moving from simple agent automations to a full cloud software factory. Start with things like PR reviews, bug triage, and CI fixes, then connect them into an end-to-end loop across spec, implementation, review, verification, and monitoring. Once that works on a simple app, you can scale it with shared context, evals, governance, and multi-model routing.

Developer Matt Pocock shipped an update to his /retro skill that turns coding-agent mistakes into permanent guardrails. Instead of adding every fix to CODING_STANDARDS.md and hoping the agent remembers, it converts fixable patterns into deterministic checks.

The skill classifies each finding. Mechanical violations (banned APIs, import shapes, file-location rules) become lint rules, pre-commit hooks, or CI jobs. Judgement calls stay in CODING_STANDARDS.md for the reviewer agent. A repo with no linting or CI at all now gets flagged as a finding on its own.

Pro tip: if the install fails with 'plugin not found,' run claude plugins marketplace update

P.S. Get 50+ AI coding hacks for Claude Code, Cursor, and Codex here.

A dev gave his AI agent $100 and told it to fund itself for another month. Its money-making strategy left him horrified.

Grow customers & revenue: Join companies like Google, IBM, and Datadog. Showcase your product to our 350K+ engineers and 150K+ followers on socials. Get in touch.

We put together a few guides on coding agents, agentic engineering, and leadership frameworks to help you level up in your career. Browse all our guides.

You can also reply directly to this email if you have suggestions, feedback, or questions.

Until next time — The Code team